1 00:00:00,000 --> 00:00:12,200 Welcome to cannabis data science. You're in for a particular treat today. You know, as 2 00:00:12,200 --> 00:00:20,960 we've been at this for a long time now, just slow and methodically collecting data, standardizing 3 00:00:20,960 --> 00:00:30,320 it, looking at it, asking interesting research questions, trying to draw insights. And we try 4 00:00:30,320 --> 00:00:37,120 to follow the demand. We try to listen, keep our ears to the ground, hear what statistics are 5 00:00:37,120 --> 00:00:43,240 people demanding, and then try to put those statistics and data into people's hands because 6 00:00:43,240 --> 00:00:49,400 we think that's our comparative advantage and the way that we can help people out in the cannabis 7 00:00:49,400 --> 00:00:59,040 space. I'll share with you a lot of new ground, a lot of work that can be done, some exciting new 8 00:00:59,040 --> 00:01:04,240 developments, and of course, we'll just have a fun day crunching cannabis data at the end of the day. 9 00:01:04,240 --> 00:01:09,960 So that's what I have to share with you today. We'd love to hear about some of your adventures. 10 00:01:09,960 --> 00:01:16,880 It's a meetup after all, so you know, feel free to share some of your interests. So I'll start, 11 00:01:16,880 --> 00:01:24,760 so I don't put anyone on the spot. I'll start with my intrepid co-host, Candice, who's been 12 00:01:24,760 --> 00:01:32,120 helping me wrangle data for a long time now. She's got an impressive technical setup. So I'll pass 13 00:01:32,120 --> 00:01:37,960 it off to you, Candice, anything you'd like to share? Any of your thoughts you want to put on 14 00:01:37,960 --> 00:01:46,840 the table? Well, it's been a fun week. I ran Get Results or Get Licenses for Florida. That was 15 00:01:46,840 --> 00:01:54,400 pretty cool. Came up with real nice CSV. Thank you, Keegan. And also to Get Results Florida, 16 00:01:54,400 --> 00:02:04,720 I'm getting those as well. I have 17.3 gigabyte of COAPDS, and actually, I made some changes in the 17 00:02:04,720 --> 00:02:11,960 code that where I resulted with Cannot Save File into a Non-Existent Directory Parent, but I know 18 00:02:11,960 --> 00:02:20,840 where that is. And otherwise, yeah, it's, you know, parsing and putting into a results data set. And 19 00:02:20,840 --> 00:02:27,640 I'm really excited because it's a lot of information for Florida. And I have been trying to obtain it by 20 00:02:27,640 --> 00:02:35,440 purchasing product and, you know, with Facebook group medical patients in Florida that were 21 00:02:35,440 --> 00:02:44,640 sharing COAs. So I'm really excited about this. And that and the metric API, I'm getting closer to 22 00:02:44,640 --> 00:02:52,120 that. I'm just waiting now. I filled out everything, the agreement. And as far as the Massachusetts 23 00:02:52,120 --> 00:02:59,320 Freedom of Information Act, I am going to pursue it. You know, because, you know, like, CCC does 24 00:02:59,320 --> 00:03:05,200 have the records, they have it in a database, and all they have to do is just publish it to a CSV. 25 00:03:05,200 --> 00:03:11,800 file instead of printing it out. But they're saying that they don't have the records. And I don't 26 00:03:11,800 --> 00:03:18,160 know, you know, I, I could be mistaken, right. But I just want to ask them, you know, about that. I 27 00:03:18,160 --> 00:03:24,520 really don't understand why it can't be done, or it can't be done by other states. But that's it. 28 00:03:25,920 --> 00:03:28,040 And hi, everybody. 29 00:03:28,040 --> 00:03:38,120 I absolutely love it. Because that's sort of the lesson of the day is ask and you shall receive. 30 00:03:38,120 --> 00:03:50,480 And just the act of asking, as we've discussed in the past, and I'll drill home today, adds value, 31 00:03:51,040 --> 00:03:56,880 right. And I always say, right, one of the lessons in the past was, you know, it doesn't hurt to ask. 32 00:03:56,880 --> 00:04:03,560 So, as Candice was saying, you know, she's at least asking for lab results for some of these cannabis 33 00:04:03,560 --> 00:04:11,400 products in her state, because those are data points that she's interested in. And Candice isn't 34 00:04:11,400 --> 00:04:18,840 alone. There's a lot of people out there searching for this data. I'll actually piggyback on a lot of 35 00:04:18,840 --> 00:04:26,920 the things Candice said momentarily. But I'll have a lot to say on that. So while I'm, you know, 36 00:04:26,920 --> 00:04:32,680 formalizing my thoughts, I'll go ahead and let everybody have a chance to speak. So, it's 37 00:04:32,680 --> 00:04:41,320 phenomenal to see you. You've been on my mind as we've got some new data science algorithms to 38 00:04:41,320 --> 00:04:49,120 write, and Canlytics is going to be looking for some of its first hires in the near future. So, 39 00:04:49,120 --> 00:04:57,640 I'm going to be getting everybody formally contacted once I get all the job postings listed. As I will 40 00:04:57,640 --> 00:05:08,800 share with you today, I've been sprinting on something interesting. I can go back to just a nice 41 00:05:08,800 --> 00:05:20,480 methodical jog at this point, and try to get some of you awesome people involved. So, but what's on 42 00:05:20,480 --> 00:05:26,240 your mind or plate or any thoughts you want to put on the table? I'm doing a data science boot camp 43 00:05:26,240 --> 00:05:32,040 right now. Well, hi, my name is Hector. I sporadically show up in these meetings every once in a while. 44 00:05:32,040 --> 00:05:37,960 I'm doing a data science boot camp. And right now we're on break, we just finished our second capstone. 45 00:05:37,960 --> 00:05:53,440 And I did a project on identifying candidates who were going to go bankrupt. And so I ran several 46 00:05:53,440 --> 00:06:03,440 machine learning algorithms, including random forest and XGBoost. And I was able to get some 47 00:06:03,440 --> 00:06:12,640 pretty good results from running my models. I ran PCA, I did a whole bunch of stuff. It was a lot 48 00:06:12,640 --> 00:06:20,480 of fun. I look forward to being able to use my skills even further in the future. Well, in the 49 00:06:20,480 --> 00:06:27,960 past, remember, we poked around at survival analysis. I think we were looking at Washington 50 00:06:27,960 --> 00:06:35,120 State, but we've got traceability data, and even lab results, it contains a surprising amount of 51 00:06:35,120 --> 00:06:44,640 information. And one thing is just who's operating. So just from say lab results alone, you can kind of 52 00:06:44,640 --> 00:06:53,200 get a pulse of who's going to the lab, how many samples are they sending in. But in Washington 53 00:06:53,200 --> 00:06:59,520 State, we've got full traceability data, so we can even look at sales. And so we were saying, Oh, you 54 00:06:59,520 --> 00:07:04,920 know, we're, we're trying to look at people's sales over time. And we noticed people would drop out of 55 00:07:04,920 --> 00:07:10,840 the market. And of course, people would enter the market. So what you could potentially do is maybe 56 00:07:10,840 --> 00:07:18,280 you could adapt your bankruptcy model to the Washington State data, and almost have a risk 57 00:07:18,280 --> 00:07:25,400 assessment model, or slash survival model. So maybe you could apply it to licensees there. And yeah, 58 00:07:25,400 --> 00:07:34,360 maybe some of the licensees, they're going along steady, they're at low risk of bankruptcy, or, or 59 00:07:34,360 --> 00:07:42,000 falling out of the market. But maybe you can identify people who are at high risk of exiting, 60 00:07:42,000 --> 00:07:49,600 and they may not even realize it. So you may be able to send them an email and just say, Hey, some 61 00:07:49,600 --> 00:07:57,160 of your key performance indicators aren't looking that great, or I've got this risk assessment model, 62 00:07:57,160 --> 00:08:04,400 and you know, maybe you were flagged. And, you know, maybe we could, you know, work together to try to 63 00:08:04,400 --> 00:08:15,840 figure out some ways to, to, to lower your risk of exit. So, but I'm just brainstorming, I'm sure you 64 00:08:15,840 --> 00:08:25,880 can think of, you know, many more uses or even a better, better application than that. But I love 65 00:08:25,880 --> 00:08:30,400 that you're, you're keeping your nose to the grindstone and working on your skills. That's important. 66 00:08:30,400 --> 00:08:43,760 So, Robert, good to see you. I think we were corresponding about some, some work to be done. And 67 00:08:43,760 --> 00:08:50,080 I've started to share some of the, the data collection efforts, slash standardization efforts, 68 00:08:50,080 --> 00:08:55,880 even the AI efforts that we're working on in Slack, and so on and so forth. So please, please, 69 00:08:55,880 --> 00:09:01,160 everyone get in touch with me if you need an invite to the Slack channel or anything. But how 70 00:09:01,160 --> 00:09:10,400 is your work going, Robert? Good, yes, similar to Hector. I'm also in a data science boot camp and 71 00:09:10,400 --> 00:09:18,320 finished up Capstone 2 projects where I did something somewhat similar. And that was looking 72 00:09:18,320 --> 00:09:30,560 at a credit card, credit scoring classification model where I used about six different classification 73 00:09:30,560 --> 00:09:40,520 algorithms. And then before that, I also had a fair amount of data scrubbing and standardizing 74 00:09:40,520 --> 00:09:49,520 to do. And, you know, most, or in the past, as a SQL developer, I would, you know, not be crazy 75 00:09:49,520 --> 00:09:56,360 about the idea of data scrubbing, but it was a good chance to practice Python. So I was pretty 76 00:09:56,360 --> 00:10:06,720 happy or totally fine with running a bunch of different Python scripts to get clean CSV files 77 00:10:06,720 --> 00:10:15,240 that would be able to import into a SQL database without generating errors and so on, import errors. 78 00:10:15,240 --> 00:10:24,200 And then similarly, I'm, you know, these days, super enthused with, you know, what AI and chat 79 00:10:24,200 --> 00:10:33,640 GPT and what Bard can do. Very interesting stuff going on right now with all this AI that's out 80 00:10:33,640 --> 00:10:41,040 there. So yeah, that's me, you know, studying hard and I'll circle back to you on getting some of 81 00:10:41,040 --> 00:10:51,680 those data files and how I can assist. Well, once again, Robert, going to tie in your interests as 82 00:10:51,680 --> 00:10:59,440 well, because I always like to explore all the latest and greatest tools. I always caution people, 83 00:10:59,440 --> 00:11:05,760 you know, don't put all your eggs in, in one basket. You know, in the past, we talked about, 84 00:11:05,760 --> 00:11:11,520 you know, some of the blockchain technologies. And while I still think those are awesome 85 00:11:11,520 --> 00:11:18,880 technologies, and to a certain extent, we use little pieces of them, we use hashing. And as I 86 00:11:18,880 --> 00:11:23,280 was talking about, like there's ways that we can start incorporating more of those ideas into our 87 00:11:23,280 --> 00:11:32,320 work. But I think the slow approach is good, just slow adoption. As I said, don't just want to rush 88 00:11:32,320 --> 00:11:39,440 full steam into it. So I'm a little careful with AI in that I don't necessarily want to rush full 89 00:11:39,440 --> 00:11:46,400 steam into it. But at the same time, I think it would be negligent to not explore it. It's a 90 00:11:46,400 --> 00:11:52,720 awesome new technology, everyone's raving about it. So shouldn't we understand it fairly well? 91 00:11:53,360 --> 00:12:04,720 And that's truly, as I'll show you today, it can be interesting to help us do well estimation. 92 00:12:04,720 --> 00:12:12,480 Remember, statistics at the end of the day is about prediction. So today, I'll show you how we can 93 00:12:12,480 --> 00:12:20,400 use the AI model for predictive purposes. So we may not want to say that this is the definitive 94 00:12:20,400 --> 00:12:28,320 answer. But sometimes, remember, when we were doing forecasting, I was saying, you know, any 95 00:12:28,320 --> 00:12:35,760 forecast is better than no forecast. And so in this case, it will be any look at the data is better 96 00:12:35,760 --> 00:12:42,960 than no look at the data. So that's a teaser to really, really exciting things to share with you 97 00:12:42,960 --> 00:12:51,520 momentarily. But before I get to that, Rick, thrilled to see you today. Going to be picking 98 00:12:51,520 --> 00:12:57,760 up with some of the topics from last week, in particular, Dino hunting. So that's sort of the 99 00:12:57,760 --> 00:13:02,640 the theme of the day today. So we'd love to hear about some of your thoughts, or ideas that you 100 00:13:02,640 --> 00:13:09,040 want to put on the table. Yeah, I'm excited to see what you've come up with for today. It was a 101 00:13:09,040 --> 00:13:17,280 great conversation last week. So hi, everyone. My name is Rick. Right now, my main focus is AI. I've 102 00:13:17,280 --> 00:13:26,320 been working not so much with collecting number data, but more text data. This is I don't know if 103 00:13:26,320 --> 00:13:32,480 you consider it like tribal knowledge or what, but in terms of cultivating, I think it's a 104 00:13:32,480 --> 00:13:40,720 great idea to be collecting, you know, history about genetics and where they came from and trying 105 00:13:40,720 --> 00:13:48,400 to track origins of specific expressions and plants, so on and so forth. There's not a huge 106 00:13:48,400 --> 00:13:54,240 database available for collecting a lot of that there is books and stuff. So I've essentially 107 00:13:54,240 --> 00:14:00,240 just been scraping all of that data from various forums like Reddit, there's a lot of good 108 00:14:00,240 --> 00:14:07,920 things there. And then running those through an embedding process and then storing it in a vector 109 00:14:07,920 --> 00:14:15,920 database for retrieval for AI. So you just kind of throw like an AI chat bot on top of that and it 110 00:14:15,920 --> 00:14:22,960 can query the database and get really accurate responses after, you know, a minimal amount of 111 00:14:22,960 --> 00:14:29,680 fine tuning. It's got a lot of moving pieces and it's hard to keep up with because things change 112 00:14:29,680 --> 00:14:35,920 and I haven't used it for numbers yet. Gigan and I suspect it's because of a lot of your cautions 113 00:14:35,920 --> 00:14:41,440 as well. But for text and stuff like growing or helping someone that has a cultivation question, 114 00:14:41,440 --> 00:14:45,440 it's working pretty well right now. So that's what I'm working on. 115 00:14:46,960 --> 00:14:54,640 I absolutely love it, Rick, and hopefully we can add some tools to your tool belt today and, you 116 00:14:54,640 --> 00:15:01,280 know, just kind of, you know, bounce good ideas off of each other because a lot of the work we're 117 00:15:01,280 --> 00:15:13,200 doing is complementary. So absolutely love it. Excited to see, you know, how you develop 118 00:15:13,200 --> 00:15:18,240 and how your projects grow. So keep us up to date on these. So this is exciting work that you're doing. 119 00:15:18,240 --> 00:15:24,480 It's right in the vein of everything we're doing today. And it's funny that you once again bring 120 00:15:24,480 --> 00:15:31,680 up Reddit because I'll tie in some of the work you're doing today too. So simply phenomenal. 121 00:15:31,680 --> 00:15:38,240 Simply phenomenal. How about you, Anna? Welcome to the group. Well, you've been to the group. So 122 00:15:38,240 --> 00:15:43,840 good to see you today. Thrilled to see you. We'd love to hear about some of the things that are 123 00:15:43,840 --> 00:15:48,800 on your plate and any ideas that you want to put on the table. Oh my gosh. I haven't been working 124 00:15:48,800 --> 00:15:53,440 on anything exciting. I've just been doing a lot of housekeeping. So setting up the 125 00:15:54,720 --> 00:16:02,800 external hard drive and just my file system and all that. But I am finding that there's not a whole 126 00:16:02,800 --> 00:16:10,240 lot of data for Oregon to scrape off the web. And so I've still been looking around. I've 127 00:16:10,240 --> 00:16:17,200 guess, yeah, I'm still a little bit lost on that part. And I am looking for other people from 128 00:16:17,200 --> 00:16:24,240 Portland or Oregon in general. So yeah, I may be shifting off of the medical focus, even though 129 00:16:24,240 --> 00:16:29,520 that's always where I want to put my efforts. But right now I'm a little distracted with some 130 00:16:29,520 --> 00:16:36,240 of the stuff going on in the policy arena. So yeah, but I don't really have anything very 131 00:16:36,240 --> 00:16:41,760 interesting to present just puttering along still. But I just wanted to reconnect for sure. So 132 00:16:42,400 --> 00:16:48,400 thank you. I'm in good company, it looks like. Well, Anna, I think we can give you some rays of 133 00:16:48,400 --> 00:17:00,480 sunshine here because I think we can provide or you can maybe even Kindle, Spark up a way to 134 00:17:00,480 --> 00:17:09,600 actually get some of this Oregon data. And once again, it all begins, as we said, with asking. 135 00:17:09,600 --> 00:17:16,480 So I'll go ahead and share my screen here, but you're definitely on the right path. And as I 136 00:17:16,480 --> 00:17:27,120 tell everybody, you know, even if you're just able to spend 20 minutes a day, just doing some sort of 137 00:17:27,120 --> 00:17:33,440 data science thing, right, even if that's just setting up, like you said, your external hard drive, 138 00:17:33,440 --> 00:17:42,640 that's still moving the ball forward. And these external hard drives, I've gone, this is my third 139 00:17:42,640 --> 00:17:49,360 one now. So be really careful with them. And because they're really, they're actually kind of 140 00:17:49,360 --> 00:17:57,520 fragile. And so like I did, you drop it on the floor, it's, it's gone, I'm going to try to do like 141 00:17:57,520 --> 00:18:06,320 some some data recovery. Once, once again, my philosophy is, you know, archive, archive, archive. 142 00:18:06,320 --> 00:18:12,720 So you want to have data stored in multiple places. And that's why, yes, you know, it's phenomenal 143 00:18:12,720 --> 00:18:19,280 for me to have it on my external hard drive. And I think that's a great way to do that. And 144 00:18:19,280 --> 00:18:25,280 you know, if you can afford it, to put it in the cloud, you can do that. But you know, cloud prices 145 00:18:25,280 --> 00:18:37,280 can be non negligible. And then distribute, that's the our third source of resiliency. So if I drop 146 00:18:37,280 --> 00:18:45,200 my hard drive with a bunch of this lab results, you know, on the floor, like I did, that's okay, 147 00:18:45,200 --> 00:18:53,200 because all of you awesome data scientists are also archiving data. So you know, as Candice has 148 00:18:53,200 --> 00:19:00,640 mentioned, she's beginning her archive of lab results. And I would encourage all of you to, to 149 00:19:00,640 --> 00:19:06,720 archive data, because just, you know, maybe I dropped my hard drive on the floor, Candice drops 150 00:19:06,720 --> 00:19:15,200 her hard drive on the floor. But, you know, thankfully, Anna has has a copy. So, you know, you 151 00:19:15,200 --> 00:19:22,400 just never know when your copy is going to pay off or be useful to someone. And Anna, you also 152 00:19:22,400 --> 00:19:33,680 mentioned medical patients. Well, as Candice was mentioning in the beginning, in Florida, there's 153 00:19:33,680 --> 00:19:45,120 medical cannabis, and look at this. So we were telling people that, you know, it doesn't hurt to 154 00:19:45,120 --> 00:19:52,880 ask, you know, you're going to be going to a dispensary, I think, you know, that if you're 155 00:19:52,880 --> 00:19:59,920 consuming a product, it's pertinent to know what you're consuming. So it wouldn't hurt to ask for 156 00:19:59,920 --> 00:20:06,960 the certificate. I don't think in Florida, they're mandated to provide them. In some states, they are 157 00:20:06,960 --> 00:20:15,600 like Washington State. And in some states, they're maybe not mandated, but people will provide them. 158 00:20:15,600 --> 00:20:21,760 So I've seen them in Massachusetts before from just asking at a dispensary, and I don't think 159 00:20:21,760 --> 00:20:31,920 it's mandated there. But this is pretty typical, right? And as I said, don't get mad, right? So 160 00:20:34,640 --> 00:20:40,400 just because it's okay to ask doesn't necessarily mean they have to say yes. So, you know, you may 161 00:20:40,400 --> 00:20:47,360 get a no, and you have to kind of just be okay with that. So here they say, oh, you know, they said, 162 00:20:47,360 --> 00:20:55,760 oh, the printer was offline. And once again, that may be perfectly the that may actually be the case 163 00:20:55,760 --> 00:21:07,040 because printers are notoriously fickle. But this is my point is, if you go around asking, well, 164 00:21:07,040 --> 00:21:16,800 so that was three months ago. Well, eight days ago, look at this, somebody made a post, you know, 165 00:21:16,800 --> 00:21:22,880 Dita can finally put their CEO is up. And so they said they thought, oh, you know, they never thought 166 00:21:22,880 --> 00:21:30,400 it would happen. But they're like, oh, this is amazing. You know, I can now shop off of turpies. 167 00:21:32,640 --> 00:21:39,360 And in like, this is another thing is, you know, you shouldn't have to look at the COA 168 00:21:39,360 --> 00:21:46,160 after you purchased it, right? Because ideally, you want to know what you're consuming, 169 00:21:46,160 --> 00:21:56,000 you know, before you purchase it. And so I'll let you explore this. But, you know, now people 170 00:21:56,000 --> 00:22:03,680 are saying like, oh, look, you know, here, I bought this product. And they're saying, oh, you know, 171 00:22:04,480 --> 00:22:12,880 what's up with the THC values? And then, you know, people chime in, in the comments, you know, kind of, 172 00:22:12,880 --> 00:22:19,600 you know, having a conversation about it. And I think this is this is brilliant, right? Because 173 00:22:20,320 --> 00:22:27,760 this is kind of reducing some of the the murkiness that was in the market, right? People were 174 00:22:27,760 --> 00:22:33,680 kind of, you know, getting hung up on these THC numbers. But now they're actually kind of able to 175 00:22:33,680 --> 00:22:44,000 see like, oh, like, there's more to it. Then, you know, there's more to this than just total THC. 176 00:22:44,000 --> 00:22:50,160 You know, we've got all the the other cannabinoids. And oh, you know, what's going on, you know, here 177 00:22:50,160 --> 00:23:01,600 with with turpines? And sure enough, so here's somebody saying, like, oh, like, so I think I like 178 00:23:01,600 --> 00:23:08,640 terpenes, right? So and they're saying, you know, oh, you know, what's the deal with with barnesine, 179 00:23:08,640 --> 00:23:16,480 which, which is a terpene they test for in in Florida. And so once again, you know, this person 180 00:23:16,480 --> 00:23:23,440 just finally, you know, saw some COAs, and they're just asking, like, okay, like, they didn't even 181 00:23:23,440 --> 00:23:29,120 know about terpenes. And they're like, you know, can someone just please explain terpenes to me? 182 00:23:29,120 --> 00:23:35,440 And, you know, sure enough, you know, once again, you've got to be cautious about things people tell 183 00:23:35,440 --> 00:23:42,800 you, right, you want to, you know, check it for factual accuracy and everything. But this is still 184 00:23:42,800 --> 00:23:51,680 a conversation. And so that's the point is, ask a question, and it can, you know, get the conversation 185 00:23:51,680 --> 00:24:01,600 going. They asked the question about COAs got the conversation going. You know, they posted a COA, 186 00:24:02,480 --> 00:24:09,360 you know, that got the conversation going to terpenes. And then, you know, and then now, Rick, 187 00:24:09,360 --> 00:24:16,880 now this is sort of up your vein. People are now saying like, okay, you know, is there any way we 188 00:24:16,880 --> 00:24:23,440 can start, you know, accumulating some of these COAs, so that way we can kind of understand them 189 00:24:23,440 --> 00:24:35,600 as a whole. So for example, if you see beta pionene at 0.2%, is that a lot? Or is that a little? Well, 190 00:24:35,600 --> 00:24:41,280 as we've seen, if you actually look at the distribution, you can find out if something 191 00:24:41,280 --> 00:24:48,560 is, you know, above average or below average, and that's the useful data point. So that's what 192 00:24:48,560 --> 00:24:58,080 people are kind of after. And once again, the Rick, these aren't finished projects. And so this is 193 00:24:58,080 --> 00:25:05,280 what I call a demand for a product, right? So these are people saying that, oh, you know, we want 194 00:25:05,280 --> 00:25:14,880 some COAs, you know, we want a COA database. Here's somebody who's got a research question. So they say, 195 00:25:14,880 --> 00:25:23,440 you know, oh, you know, what's going on with with Farnes? And now finally, you know, somebody just 196 00:25:23,440 --> 00:25:31,680 asks the question, you know, can we just get a, you know, a universal COA search site? Well, 197 00:25:31,680 --> 00:25:43,440 remember at the beginning, ask, and you shall receive. So this is, you know, just the Reddit, 198 00:25:43,440 --> 00:25:51,360 on Reddit was just the crude origins, you know, to this, you know, very, very day, or a couple days 199 00:25:51,360 --> 00:25:59,120 ago, you know, people are just posting up their results. And, you know, you can get the the various 200 00:25:59,120 --> 00:26:08,640 you know, you can start looking at some of these COAs. And, you know, so what I realized is, oh, 201 00:26:08,640 --> 00:26:18,160 well, they have, you know, QR codes on them. So what you can do is scan the QR code on your URL, 202 00:26:18,160 --> 00:26:25,520 or I mean, scan the QR code on your phone, or with with Python, or however you please, 203 00:26:25,520 --> 00:26:35,600 and you know, you can now get this. And actually, here, we can actually do this in Python real quick. 204 00:26:35,600 --> 00:26:42,800 So to just show you how you can actually get the COA URL off of this thing. And, you know, 205 00:26:42,800 --> 00:26:50,880 you can use CANlytics. And our favorite COA parsing tool, and you can use the 206 00:26:50,880 --> 00:27:04,400 COA doc. And then we can find the COA URL over here. So I think we can do this without too much 207 00:27:04,400 --> 00:27:11,520 trouble. Hopefully, if it's too much trouble, I'll just pass it. But we can just say, oh, let's find 208 00:27:11,520 --> 00:27:19,440 the COA URL. And then we can find the COA URL over here. So I think we can do this without too much 209 00:27:19,440 --> 00:27:29,840 trouble. And then we can just say, oh, let's find this QR code. And so this just checks all the images 210 00:27:29,840 --> 00:27:39,200 and sees if any of them are QR codes, and if there's a parsable URL from them. And so once again, 211 00:27:39,200 --> 00:27:46,000 not super fast. So if any of you want to get in there, and also I'm running a bunch of processes 212 00:27:46,000 --> 00:27:58,400 now. But long story short, you know, you can start to, you know, find the direct source of these 213 00:27:59,600 --> 00:28:10,960 COAs. And basically, upon, you know, further exploration, you know, you can basically find 214 00:28:10,960 --> 00:28:24,080 that these various laboratories are publishing, I think, similar to NCR labs, where maybe certain 215 00:28:24,080 --> 00:28:32,240 clients want to make their COAs publicly available. So for example, you know, VitaCAN, 216 00:28:32,240 --> 00:28:46,640 you know, they're trying to get their COAs in the hands of consumers. And just to kind of show you, 217 00:28:48,000 --> 00:28:54,640 so for example, one of the things like Candice was working on, that I was working on, was just 218 00:28:54,640 --> 00:29:12,400 archiving all of these. So basically, let's see if we can't, here one second here, just going to see 219 00:29:12,400 --> 00:29:21,600 if we can't just pull these up real quick for you. So for example, say we wanted to look at VitaCAN 220 00:29:21,600 --> 00:29:33,360 to look at VitaCAN COAs. We can do that. Okay, let me try one more thing. 221 00:29:33,360 --> 00:29:49,760 Okay, maybe not. I think I just had an extra backslash in there. Okay, let's try one more time. 222 00:29:52,160 --> 00:30:01,840 Okay, so they just have, so remember, they just started, you know, eight days ago or so. So they 223 00:30:01,840 --> 00:30:10,240 may not have, well, actually these look like they're from earlier. Oh, so here's the latest one. 224 00:30:10,240 --> 00:30:20,160 So here's one from April 28th. So maybe they don't have all of their COAs here, but you can, 225 00:30:20,160 --> 00:30:32,800 you know, start to look at them. And once again, they've got the terpene values. 226 00:30:32,800 --> 00:30:39,680 And so, okay, so why am I jammering on about all of this? 227 00:30:39,680 --> 00:30:58,880 Well, there's a demand for COA data. So, you know, why don't we basically, you know, 228 00:30:58,880 --> 00:31:05,440 give the people what they want. So, you know, what you can do is, you know, 229 00:31:05,440 --> 00:31:16,720 you can at least find all of these COA links. So, you know, there's, we were just doing them, 230 00:31:17,280 --> 00:31:24,400 finding them through crude manners. So, you know, people, so here's someone who bought some animal 231 00:31:24,400 --> 00:31:37,520 tsunami. And sure enough, they posted their COA. You know, you can basically get the QR code, 232 00:31:38,960 --> 00:31:48,240 then get the official COA. That's my recommendation, because as you can see, people's pictures 233 00:31:48,240 --> 00:31:57,600 sometimes cut off portions of the certificate. So it's best if you can get the official URL, 234 00:31:58,640 --> 00:32:05,760 you know, as we said, you know, go straight to the source. And then, you know, and then we can 235 00:32:05,760 --> 00:32:13,600 basically work on COA parsing tools, because that's where people are around the world. 236 00:32:13,600 --> 00:32:20,080 COA parsing tools, because that's what people are worried about, right? They're saying like, 237 00:32:20,080 --> 00:32:28,800 oh, you know, what's going on with barnesine? And, you know, we've been interested in 238 00:32:28,800 --> 00:32:37,760 limonene and, you know, beta pinyin. And it doesn't really make much sense to just 239 00:32:37,760 --> 00:32:49,200 have to to plough through this COA. So, you know, why don't we basically do this with Python here? 240 00:32:50,000 --> 00:32:56,800 So I'll go ahead and show you how we can do that. But before I do that, any thoughts, comments, 241 00:32:56,800 --> 00:33:08,160 questions before we dive into this code here? Okay. Still with you all, right? 242 00:33:10,560 --> 00:33:19,840 Yep. Okay, phenomenal. So I'll try to tie this, tie this all together and bring it home. So long 243 00:33:19,840 --> 00:33:30,560 story short, you can start archiving some of these COAs. And this is just a little setup here. So, 244 00:33:30,560 --> 00:33:41,280 as I said, we're going to use OpenAIs. So what we can do, well, right, we can find this PDF. 245 00:33:41,280 --> 00:33:50,160 And so, right, so people are posting up their various PDFs here. And so this, this is just 246 00:33:50,720 --> 00:34:03,440 a sample here that somebody tested this. See, it's a goo berry extraction. And this one was tested at 247 00:34:03,440 --> 00:34:13,040 ACS laboratory. And so here's one, right, this one was at Keisha Labs. And I am working on it, 248 00:34:13,040 --> 00:34:22,880 basically a parsing algorithm for for some of these, you know, Keisha certificates, and some of 249 00:34:22,880 --> 00:34:32,320 these ACS certificates. But I started looking at some of this, some of the data, some of these 250 00:34:32,320 --> 00:34:42,000 certificates. So I'll just show you some of the PDFs here. Here, I'll try to find one that's got 251 00:34:42,000 --> 00:35:02,640 a large span of... So, right, here's, you know, like a more recent COA, a one to one CBD THC ball. 252 00:35:04,000 --> 00:35:11,040 Interesting, you know, you can, you know, you're not going to see many terpenes in this ball. 253 00:35:11,040 --> 00:35:25,200 But for some reason, they put in put in this one, a hexa, what's it say, hexa drothymol. So interesting. 254 00:35:26,480 --> 00:35:32,800 But remember, this is a ball. The important thing on this one is just the CBD and the THC 255 00:35:32,800 --> 00:35:43,600 concentrations. And but I realized, you know, not all these COAs are made the same. And, you know, 256 00:35:43,600 --> 00:35:56,400 if you go way back in time, so here's a COA from either... So this looks like a COA from 2019. 257 00:35:56,400 --> 00:36:05,920 And, you know, you don't have as much information. And in fact, you don't even have total THC or CBD 258 00:36:05,920 --> 00:36:14,320 on this one. And as you can kind of see from like the file sizes here, like over time, 259 00:36:15,760 --> 00:36:24,240 there's COA just, you know, gradually changed into, you know, what it is today. But some of the, 260 00:36:24,240 --> 00:36:34,000 you know, the earlier ones are going to be a little different, right? So here's like an early 261 00:36:34,560 --> 00:36:44,880 Caesia lab COA, where this one was in 2019. This one does have total cannabinoids on it, 262 00:36:44,880 --> 00:36:52,800 which is what we're after, at least for some of our analyses, we would like terpenes too, 263 00:36:52,800 --> 00:36:58,560 but at the minimum, we'd like to get cannabinoids. And so what I realized is, you know, it may not 264 00:36:58,560 --> 00:37:11,520 actually make sense to write an algorithm to parse a particular COA, because the formatting's always 265 00:37:11,520 --> 00:37:17,840 changing. And, you know, why is it changing? Well, because, you know, software developers like to be 266 00:37:17,840 --> 00:37:24,880 continuously improving upon their code. And, you know, the laboratories would like to be, you know, 267 00:37:25,600 --> 00:37:31,440 hopefully continuously improving their COAs, you know, to a certain extent. And so, you know, 268 00:37:31,440 --> 00:37:37,280 they maybe they looked at this and they realized, oh, you know, we could add terpenes to it. And so 269 00:37:37,280 --> 00:37:46,720 just over time, you can see, you know, more styling was done, various elements moved around on the 270 00:37:46,720 --> 00:37:55,920 pages, which is, you know, fine and dandy, but it's going to make COA parsing super difficult. But 271 00:37:56,640 --> 00:38:03,680 why should we let that stop us? Because, you know, at the end of the day, the data's there, 272 00:38:03,680 --> 00:38:11,040 so remember, you know, never throw away any data. So this is sort of the problem we're faced with. 273 00:38:12,080 --> 00:38:21,600 People want these lab results. They're there. They're right there on the web. The data is even 274 00:38:21,600 --> 00:38:29,920 on the PDF. But as we said, it's basically locked and unobtainable. Like as we said, 275 00:38:29,920 --> 00:38:37,280 the data is so close, yet so far away, right? You can get all the certificates and, you know, 276 00:38:37,280 --> 00:38:43,360 people are diligently doing that here on Reddit, right? They recognize that the data is so close, 277 00:38:43,360 --> 00:38:50,880 it's so close, let's at least get these URLs. And, you know, they did a noble thing there, 278 00:38:50,880 --> 00:38:56,160 because it pointed us in the direction of them. And, you know, we've got some interesting 279 00:38:56,160 --> 00:39:09,120 tools up our tool belt here. So without further anticipation, let's just go ahead and get into it. 280 00:39:09,120 --> 00:39:18,480 So we've got the first page here of the certificate. So here's this crew clear syringe. 281 00:39:18,480 --> 00:39:32,880 So remember, this is the ACS laboratory certificate for the true clear syringe. 282 00:39:33,760 --> 00:39:43,920 Okay, so let's keep this in mind. So we basically just have all of the text from the front page. 283 00:39:43,920 --> 00:39:53,040 Let me print this out nicely for you. This is just the raw text. So there's the true clear syringe. 284 00:39:55,840 --> 00:40:01,840 You know, here's all the cannabinoid information. Oh, look, you know, there's farinezine. 285 00:40:01,840 --> 00:40:09,200 Actually, just for fun, let's just go ahead and get farinezine. Okay, we'll just let AI deal with 286 00:40:09,200 --> 00:40:26,520 the spelling on this one. So here's the 287 00:40:26,520 --> 00:40:34,880 text. So what if we just tell you, remember, let's try an untrained chat GPT model and just say, 288 00:40:34,880 --> 00:40:45,520 hey, you know, extract as many of these data points to JSON from the following text, right, 289 00:40:45,520 --> 00:40:51,520 because that's basically the same data point that we're going to extract from the following text. 290 00:40:51,520 --> 00:40:56,720 Right, because that's basically what you could tell a human to do. It would take them a long time, 291 00:40:56,720 --> 00:41:03,760 you know, a handful of minutes, and they wouldn't love to do it. But you could tell a human, okay, 292 00:41:04,560 --> 00:41:11,440 put the product name here, put the product type here, put the producer here, put the THC CBD and 293 00:41:11,440 --> 00:41:16,480 some of these terpenes here, right. And then we would go through and we would say, okay, here's 294 00:41:16,480 --> 00:41:27,440 farinezine, they want the farinezine percent, you know, it's 4.807. I'm going to give the model a 295 00:41:27,440 --> 00:41:36,560 supplement prompt here. And I basically am saying only, or I say return only JSON and always return 296 00:41:36,560 --> 00:41:44,480 at least an empty object if no data can be found. So I'm basically saying at least a few points, 297 00:41:44,480 --> 00:41:54,480 so I'm basically saying at least return me something here. And let's just see what happens. 298 00:41:54,480 --> 00:42:02,480 So there's a lot of things that can go wrong here. There's the connection error, hopefully, 299 00:42:02,480 --> 00:42:10,480 you know, I'm streaming after all. So we may not want to expect this to be super speedy, but 300 00:42:10,480 --> 00:42:18,480 as often with proof of concepts, they can always be refined. So the idea is you want to at least 301 00:42:18,480 --> 00:42:24,480 make sure you can do something and then you can, you know, do it better and faster. And so check 302 00:42:24,480 --> 00:42:34,480 this out. So we've got the product name, there's the true clear syringe. It did not get the product 303 00:42:34,480 --> 00:42:50,480 type out of this. Unless it did, but here let's look at the COA real quick. So it didn't quite get 304 00:42:50,480 --> 00:43:00,480 the sample type correct. So it's in sample matrix. Well, I don't know, that's debatable. So we'll 305 00:43:00,480 --> 00:43:10,480 maybe have to try this out on more COAs, but the product type may have been a miss. It got the 306 00:43:10,480 --> 00:43:20,480 producer as true leave. And for some reason, they put it in all capital letters, but I'm not sure why 307 00:43:20,480 --> 00:43:32,480 it did that, but we at least have true leave right there. And check this out. Total THC, we've got 87. 308 00:43:32,480 --> 00:43:38,480 Let me actually just try to put this side by side instead of going back and forth. Okay, so we're 309 00:43:38,480 --> 00:44:00,480 looking for total THC of 80. Yep. So there's 87.66. Then we've got 0.303% CBD. Then we've got 0.099% 310 00:44:00,480 --> 00:44:16,480 beta-pinene, 0.39% d-limonene, and check it out, even got the farnazine, the 4.807. And as I said, 311 00:44:16,480 --> 00:44:24,480 it would be phenomenal to get every single last data point. And so I think we should work on that. 312 00:44:24,480 --> 00:44:34,480 But I think we tried this in some prior weeks where we just tried to get the kitchen sink, and it 313 00:44:34,480 --> 00:44:40,480 didn't quite work very well. I mean, it kind of did, but it was a little hit or miss. So this is 314 00:44:40,480 --> 00:44:48,480 the GPT-4 model. So it's a slightly better model. I think last time we may have been using GPT-3 or 315 00:44:48,480 --> 00:45:02,480 3.5. So it's a slightly better model. Maybe it's a better prompt. It's a different COA. So there's a 316 00:45:02,480 --> 00:45:10,480 lot of things that we can tinker with here. So there's, as Rick was saying at the beginning, 317 00:45:10,480 --> 00:45:17,480 there's a lot of moving pieces. This is kind of an arbitrary prompt. And so this is maybe where they 318 00:45:17,480 --> 00:45:25,480 say people talk about prompt engineering. So I'll post this code. I should have done it before the 319 00:45:25,480 --> 00:45:34,480 meetup, but I'll do it immediately after. And you can try to tinker around with this. So maybe you can 320 00:45:34,480 --> 00:45:43,480 add...that would be an experiment. Just keep adding data points until maybe it breaks. So maybe it can 321 00:45:43,480 --> 00:45:51,480 handle parsing them all. But that's something for you to experiment. And there's other parameters you 322 00:45:51,480 --> 00:45:57,480 can toggle. The temperature is basically how creative you want the model to be. And I thought we 323 00:45:57,480 --> 00:46:05,480 wanted it to not be creative. We don't want it making up answers. So that could be sort of another 324 00:46:05,480 --> 00:46:16,480 check you could do. You could maybe make a query and you can say, is this number in fact in the 325 00:46:16,480 --> 00:46:30,480 COA? So you could even say, oh, is this in the front page text? And it is. So that could potentially be 326 00:46:30,480 --> 00:46:38,480 like various checks you could do after you parse it. Because remember, these are numbers. And it's 327 00:46:38,480 --> 00:46:52,480 one thing for GPT to just turn out a paragraph that you send in an email. It's another thing for it to 328 00:46:52,480 --> 00:47:02,480 turn out cannabinoid percentages that say medical patients in Florida are going to be using to make 329 00:47:02,480 --> 00:47:14,480 purchasing decisions. So we may want to put a little bit of double checking in place to be pretty 330 00:47:14,480 --> 00:47:19,480 certain about these numbers. And in fact, what people always kind of recommend with AI models is it 331 00:47:19,480 --> 00:47:25,480 doesn't hurt to have somebody check them like a human at the end of the day. So what I was thinking 332 00:47:25,480 --> 00:47:37,480 was we can work on the COA doc user interface and basically somebody could upload their certificate 333 00:47:37,480 --> 00:47:47,480 here. We could try to parse it with COA doc or OpenAI and just try to get as many data points as 334 00:47:47,480 --> 00:47:53,480 possible and then just suggest those to the user that, okay, this may be the answer. And then they 335 00:47:53,480 --> 00:48:01,480 can either say yes or no. So you could then parse the COA. You can get this list back. And they may 336 00:48:01,480 --> 00:48:09,480 say, yes, that's the product name. No, that's not the product type. Yes, that's the producer. Yes, yes, 337 00:48:09,480 --> 00:48:23,480 yes, yes. So that way you can kind of have a human confirm this at the end of the day. Well, do you 338 00:48:23,480 --> 00:48:33,480 want to see it put to use at scale? Or what are your thoughts on this one? 339 00:48:33,480 --> 00:48:39,480 Show us if you got it. I'm excited to see what you've got up your sleeve. 340 00:48:39,480 --> 00:48:48,480 Okay. Well, this one's here. I'll just get this one turning away and then let's go look at that other one. 341 00:48:48,480 --> 00:48:58,480 So basically I just wrote this into a function. So this is just the exact work we did here, but just in a 342 00:48:58,480 --> 00:49:07,480 simple function. So this is just the prompt. And so these are the things I think, right, this was just 343 00:49:07,480 --> 00:49:13,480 almost like a throwaway prompt. This is just when I kind of just came up with this morning. But these 344 00:49:13,480 --> 00:49:23,480 may be the things that people, these may be people's prized possessions. Just doing a little 345 00:49:23,480 --> 00:49:29,480 demonstration with some of the image generation, it's all about the prompts. And so people are just trying to 346 00:49:29,480 --> 00:49:35,480 figure out the right, right, that's X, right. You've got to think about this as a statistical function or 347 00:49:35,480 --> 00:49:46,480 output is Y and this is X. So this matters a lot to people. And as I said, that's what statisticians have 348 00:49:46,480 --> 00:49:58,480 fun about at lunch is talking about X, talking about what goes into the model and its parameters. 349 00:49:58,480 --> 00:50:07,480 So you can spend a lot of time on this. Okay. But anywho, just the prompt and then all this is doing is 350 00:50:07,480 --> 00:50:18,480 opening the PDF, getting the front page, formatting the prompt from the front page, then pinging open AI. 351 00:50:18,480 --> 00:50:25,480 It's either going to fail or maybe the parsing fails. But you know, we're going to try to get the data at 352 00:50:25,480 --> 00:50:32,480 the end of the day, because basically, this is what the content looks like. And then we can format that 353 00:50:32,480 --> 00:50:42,480 into a nice tidy base on. So I'll just let this one rip here with I'm just going to do five COAs because I 354 00:50:42,480 --> 00:50:53,480 had it run in full in another interpreter. And I'll basically show you the full results. Okay. 355 00:50:53,480 --> 00:51:02,480 So unfortunately, this first one failed. And this is where a lot of trial and error is needed. 356 00:51:02,480 --> 00:51:10,480 And that's why I encourage you all to to play with these prompts, because I've been kind of running this one 357 00:51:10,480 --> 00:51:22,480 on and off. And it's weird because this first observation sometimes fails and sometimes succeeds. 358 00:51:22,480 --> 00:51:32,480 Which is which is just completely strange and bizarre to me that, you know, you'd get, you know, two different 359 00:51:32,480 --> 00:51:44,480 outputs from the same query, but it happens. So I would, you know, encourage you all to try to try to figure out 360 00:51:44,480 --> 00:51:55,480 why that's happening. So that way we can fix it. But this is where I get into the what I the point I raised at the 361 00:51:55,480 --> 00:52:03,480 beginning was, this is now imperfect. You know, we're definitely in the realm of statistics now, right? 362 00:52:03,480 --> 00:52:15,480 We've got a non random sample. These are just certificates that people were posting non randomly in Florida. 363 00:52:15,480 --> 00:52:25,480 So that's our sample we're working with. It's further biased because certain queries are failing. 364 00:52:25,480 --> 00:52:34,480 So that's an unknown source of bias. So like, why are those are those queries failing? 365 00:52:34,480 --> 00:52:40,480 And it may not actually be necessarily by well, actually, I forget if this is technically bought. 366 00:52:40,480 --> 00:52:44,480 Actually, so there's different things that can happen in your statistics, right? 367 00:52:44,480 --> 00:52:56,480 So if you have measurement error or missing observations, please research if that can lead to bias or not. 368 00:52:56,480 --> 00:53:05,480 I don't know off the top of my head. I want to say that miscalculations can I forget if missing data can or not. 369 00:53:05,480 --> 00:53:12,480 I want to say it may lead to bias. So that's some statistical homework for you. 370 00:53:12,480 --> 00:53:16,480 But these are different concepts in statistics. 371 00:53:16,480 --> 00:53:24,480 You know, bias is a statistical concept, you know, missing data and miscalculations or entry error. 372 00:53:24,480 --> 00:53:29,480 Those are just characteristics of the data. 373 00:53:29,480 --> 00:53:40,480 But here you see we got four out of five. So running around 80 percent, which is awesome. 374 00:53:40,480 --> 00:53:46,480 Right. As I said at the beginning, I'd rather look at any data than no data. 375 00:53:46,480 --> 00:53:58,480 So, you know, let's look at this. So that's the first step, right, is get the data. 376 00:53:58,480 --> 00:54:02,480 Let's make sure I've got it here. 377 00:54:02,480 --> 00:54:09,480 So there it is. And now let's visualize it. 378 00:54:09,480 --> 00:54:13,480 And we've got some Khalifa Kush. 379 00:54:13,480 --> 00:54:23,480 And it's so interesting because I just it's so strange how the world works with these little coincidences. 380 00:54:23,480 --> 00:54:39,480 But I just shared a video with you all on Slack the last week about burner of cookies discussing some of the strains that he's tied to. 381 00:54:39,480 --> 00:55:00,480 And he was saying that, you know, Wiz Khalifa wanted a strain and I think they just maybe, you know, branded maybe a really potent variety of the OG Kush as, you know, Wiz Khalifa's Khalifa Kush. 382 00:55:00,480 --> 00:55:11,480 And, you know, just from what I was watching on the video, they were like saying that, like, you know, he wants to be the only one to grow this variety. 383 00:55:11,480 --> 00:55:25,480 And that's awesome. Well, as cannabis data scientists, you know, we can kind of help at least characterize some of the rates. 384 00:55:25,480 --> 00:55:32,480 So burners, the one who grew the strain, but we can, you know, help them with statistics. 385 00:55:32,480 --> 00:55:40,480 So, you know, maybe or maybe not. They may not know it. Say the average rate. 386 00:55:40,480 --> 00:55:46,480 Is this a Sativa or is this an Indica leaning strain into how much? 387 00:55:46,480 --> 00:56:00,480 And so what we can do is we can look at some of our favorite ratios here. So here, I forget how this function works. 388 00:56:00,480 --> 00:56:05,480 Just simple division. 389 00:56:05,480 --> 00:56:09,480 Okay, let's take the mean. 390 00:56:09,480 --> 00:56:25,480 So, remember, my rule of thumb is anything greater than 0.25 or anything greater than a one to four ratio of beta pinene to D-limonene. 391 00:56:25,480 --> 00:56:31,480 I would think of as a Sativa. We were looking at, say, Durban poison. 392 00:56:31,480 --> 00:56:39,480 And that's closer to a three to four or almost a one to one ratio. And that's when I think of a strong Sativa. 393 00:56:39,480 --> 00:56:45,480 So the higher this number goes, typically, I think the more Sativa leaning it is. 394 00:56:45,480 --> 00:56:50,480 And then the lower it goes, the more Indica it's leaning. 395 00:56:50,480 --> 00:56:59,480 And from looking at a sample of this data, it was looking to me like maybe it's some strong, strong Indicas. 396 00:56:59,480 --> 00:57:04,480 Maybe down towards like 0.1 or 0.15. 397 00:57:04,480 --> 00:57:16,480 So to me, this one looks like it's on the Indica side, maybe not the strongest, like the heaviest Indica out there. 398 00:57:16,480 --> 00:57:22,480 But that's how I would characterize this one. 399 00:57:22,480 --> 00:57:30,480 And then here, here I've just parsed a bunch more. So there's a whole lot more parsing to be done. 400 00:57:30,480 --> 00:57:38,480 So this is just the same script, but I just let it run on 57 strains. 401 00:57:38,480 --> 00:57:42,480 Then we can plot all of these on the map. 402 00:57:42,480 --> 00:57:49,480 So here you see these are ones that I would consider heavy Indicas. 403 00:57:49,480 --> 00:58:00,480 So this one, look at this, they're even calling this one a MPX Sativa OSFG Live ROS. 404 00:58:00,480 --> 00:58:08,480 Right. But that one is just kind of spitballing here. 405 00:58:08,480 --> 00:58:14,480 It's around 0.16. 406 00:58:14,480 --> 00:58:21,480 So it's interesting here because as we were saying, names may not always be perfect. 407 00:58:21,480 --> 00:58:38,480 But to me, this looks like it's labeled the Sativa OSFG. 408 00:58:38,480 --> 00:58:51,480 Double check this name for me, please. But this one's a Live ROS. And as I was saying, in my book, I would categorize that one as a strong Indica. 409 00:58:51,480 --> 00:58:56,480 And I would kind of, you know, different people have their different tastes. 410 00:58:56,480 --> 00:59:00,480 Some people seek out the Indicas. 411 00:59:00,480 --> 00:59:07,480 And then, you know, personally, I'm just a bit more fan of the Sativa, but everybody's got their personal preferences. 412 00:59:07,480 --> 00:59:17,480 So if I was shopping, you know, I would be looking for, you know, one of these like these yellow or green strains. 413 00:59:17,480 --> 00:59:24,480 So, you know, I'd be looking for like, you know, one of these Califa cushions. 414 00:59:24,480 --> 00:59:35,480 And so this is where I was getting into the fact, or remember we were talking about at the beginning. 415 00:59:35,480 --> 00:59:40,480 One of the things we would like to start looking at would be phenotypes. 416 00:59:40,480 --> 00:59:47,480 So what you can do here is you can just look at, hopefully. 417 00:59:47,480 --> 00:59:54,480 Yes, so you can. There's a ton of them. So let's just do like a sample of. 418 00:59:54,480 --> 01:00:02,480 Let's just look at 10 of them. 419 01:00:02,480 --> 01:00:05,480 We'll just look at, actually that may have been too many. 420 01:00:05,480 --> 01:00:14,480 So it looks like they're kind of clustering down there. Maybe it was enough. 421 01:00:14,480 --> 01:00:27,480 It looks like the Califa cushions are clustering down, you know, down here towards, as I was saying, you know, more on the Indica side. 422 01:00:27,480 --> 01:00:37,480 But it's just kind of interesting because, you know, now and again, you'll just have essentially a phenotype or phenotype. 423 01:00:37,480 --> 01:00:46,480 So this is the Califa cush strain. But just for whatever reason, just something about that plant. 424 01:00:46,480 --> 01:00:51,480 Maybe it was an environmental factor. Maybe it was a genetic factor. 425 01:00:51,480 --> 01:00:59,480 If they're growing these from seed instead of clone, if they're all clones, then it must be some sort of environmental factor. 426 01:00:59,480 --> 01:01:04,480 But something happened with these two varieties here. 427 01:01:04,480 --> 01:01:10,480 And in my book, you know, they're just more on the Sativa side. 428 01:01:10,480 --> 01:01:20,480 So maybe if, you know, say, you know, Wiz, Califa and Berner were trying to say branch this into two different product lines. 429 01:01:20,480 --> 01:01:31,480 You know, they may be able to, you know, spin off, you know, a Califa Cush Sativa variety strain. 430 01:01:31,480 --> 01:01:34,480 And they, you know, maybe they mix it with Durbin, right? 431 01:01:34,480 --> 01:01:43,480 Maybe you mix this phenotype with the Durbin, that's, you know, Sativa with some of the Califa Cush attributes that you like. 432 01:01:43,480 --> 01:01:49,480 And then maybe they just keep this phenotype around as their classic line. 433 01:01:49,480 --> 01:01:58,480 Or maybe they would say, oh, this one doesn't match the typical Califa Cush that people are expecting in the market. 434 01:01:58,480 --> 01:02:02,480 So maybe that one doesn't meet their quality control standards. 435 01:02:02,480 --> 01:02:06,480 So, you know, there's a lot you can do with this from here. 436 01:02:06,480 --> 01:02:12,480 But as I was saying, this is just a prediction, right? 437 01:02:12,480 --> 01:02:18,480 We need to actually, I've got all the code up here. 438 01:02:18,480 --> 01:02:22,480 So we've actually see this so strange, right? 439 01:02:22,480 --> 01:02:31,480 Because look, this time, right, when I just ran this code for you over here, the first AI query failed. 440 01:02:31,480 --> 01:02:42,480 But for whatever reason, when I ran it the first time, it succeeded and did give a result. 441 01:02:42,480 --> 01:02:53,480 So as I was saying, you know, more work needs to be done to figure out, you know, why are these requests succeeding and why are they failing? 442 01:02:53,480 --> 01:02:59,480 And, you know, are these results accurate? 443 01:02:59,480 --> 01:03:09,480 So, for example, you know, this one didn't report any beta-pionine or delimiting. 444 01:03:09,480 --> 01:03:17,480 And so what we can do is we can actually pull up that. 445 01:03:17,480 --> 01:03:25,480 You can actually pull up this PDF. 446 01:03:25,480 --> 01:03:34,480 And, you know, we can confirm if this thing has terpenes on it or not. 447 01:03:34,480 --> 01:03:44,480 So here we have the NC, sure enough, it's got beta-pionine and it also has delimiting. 448 01:03:44,480 --> 01:03:49,480 So this would be an inaccurate prediction. 449 01:03:49,480 --> 01:03:57,480 So actually, 0.69. 450 01:03:57,480 --> 01:04:00,480 Oh, wait, actually, hold on. 451 01:04:00,480 --> 01:04:04,480 This one actually, maybe that wasn't the null one. 452 01:04:04,480 --> 01:04:08,480 Maybe these ended up getting sorted somehow. 453 01:04:08,480 --> 01:04:18,480 Okay, so these may not be in order, but let's see if I can find one with missing values. 454 01:04:18,480 --> 01:04:23,480 Okay, so that may have actually worked out okay. 455 01:04:23,480 --> 01:04:28,480 So once again, I'm going to go ahead and start wrapping up since I've gone way over time. 456 01:04:28,480 --> 01:04:35,480 But I would encourage you all to basically, you know, we're going to have to start double checking these, right? 457 01:04:35,480 --> 01:04:41,480 Because it's like if we're going to use this model to extract cannabinoid and terpene data, 458 01:04:41,480 --> 01:04:46,480 then we want to have some measure of accuracy. 459 01:04:46,480 --> 01:05:01,480 So I still think it would be a useful exercise to say parse these through a well-defined algorithm 460 01:05:01,480 --> 01:05:11,480 that we know for certain is pulling all the terpenes and parsing them, you know, say to 99% accuracy 461 01:05:11,480 --> 01:05:15,480 and then compare that to the OpenAI model. 462 01:05:15,480 --> 01:05:20,480 And we would like this OpenAI model to, you know, approach. 463 01:05:20,480 --> 01:05:23,480 It's probably never going to be 100% accurate, 464 01:05:23,480 --> 01:05:30,480 but we would like it to at least approach 100% accuracy even if we never get there. 465 01:05:30,480 --> 01:05:44,480 So that's the first step, I think, or a next step is figure out what exactly is the accuracy of this prompt slash algorithm. 466 01:05:44,480 --> 01:05:50,480 And then, as I was saying, this takes a little bit of time here. 467 01:05:50,480 --> 01:05:58,480 So, you know, this may have taken five to 10 to 15 minutes or so to parse 50. 468 01:05:58,480 --> 01:06:05,480 And also, remember, there's a non-negligible price here. 469 01:06:05,480 --> 01:06:10,480 So actually, I've kind of raked up a little bit of a bill this morning. 470 01:06:10,480 --> 01:06:22,480 So through all of my development and parsing, so remember, you know, I parsed, you know, maybe 100 or so of these. 471 01:06:22,480 --> 01:06:26,480 It has, you know, cost me about five bucks. 472 01:06:26,480 --> 01:06:37,480 So once again, that's not the most in the world, but, you know, it's non-negligible. 473 01:06:37,480 --> 01:06:44,480 But my philosophy is these are such small costs and the data is so valuable. 474 01:06:44,480 --> 01:06:51,480 Surely, the marginal benefit exceeds these tiny marginal costs. 475 01:06:51,480 --> 01:07:00,480 So we've got north of 15,000 COAs that we could potentially parse. 476 01:07:00,480 --> 01:07:06,480 Don't have to do them all with AI, but check this out. 477 01:07:06,480 --> 01:07:16,480 I'll just tie this off at the end of the day with remember at the beginning, why are we doing all of this? 478 01:07:16,480 --> 01:07:28,480 Because somebody says, you know, can't we just get a universal COA site where we can say search and look at results? 479 01:07:28,480 --> 01:07:34,480 Well, ask and you shall receive. 480 01:07:34,480 --> 01:07:40,480 So I encourage you all to put this in your browser and smoke it. 481 01:07:40,480 --> 01:07:50,480 And this is just a work in progress, but a this is an extreme work in progress. 482 01:07:50,480 --> 01:07:54,480 So this is just a pure development version. 483 01:07:54,480 --> 01:07:59,480 And so there's still a lot more work to be done, especially on the parsing front. 484 01:07:59,480 --> 01:08:04,480 So I haven't incorporated the parsing yet, but this is just a retrieval tool. 485 01:08:04,480 --> 01:08:14,480 But, you know, now if we start populating a database, this may have been too much to ask for. 486 01:08:14,480 --> 01:08:17,480 OK, may have crashed. OK. 487 01:08:17,480 --> 01:08:25,480 So as I said, there's more work to be done to make sure that the website doesn't crash. 488 01:08:25,480 --> 01:08:39,480 And you can do cool things with this list, you know, like all the fun things people like to do, like sort, filter, search, download. 489 01:08:39,480 --> 01:08:48,480 So right now it just has the this one doesn't I'm not sure if this is a valid PDF. 490 01:08:48,480 --> 01:08:50,480 Oh, yeah. Check that out. 491 01:08:50,480 --> 01:08:56,480 So here's just a rosin from from green scientific. 492 01:08:56,480 --> 01:09:04,480 So, you know, this and the search algorithm I cooked up last night is is far from perfect. 493 01:09:04,480 --> 01:09:12,480 So there's a thousand and one ways that this can be improved upon or just taking an entirely different direction. 494 01:09:12,480 --> 01:09:19,480 So, Rick, you know, if you want to to work on, you know, an entirely different aspect, you're welcome to. 495 01:09:19,480 --> 01:09:26,480 There's just this is just a demonstration that, hey, this data is there. 496 01:09:26,480 --> 01:09:32,480 These are just links to PDFs that exist on the Web. 497 01:09:32,480 --> 01:09:37,480 We can make these readily searchable. 498 01:09:37,480 --> 01:09:40,480 And so now we can at least get these ways. 499 01:09:40,480 --> 01:09:47,480 So now somebody can at least search and find a way. 500 01:09:47,480 --> 01:09:57,480 Well, you know, now we have written just this morning an algorithm to actually parse this. 501 01:09:57,480 --> 01:10:03,480 So once again, it's going to be a non zero cost. 502 01:10:03,480 --> 01:10:14,480 But somebody could now try to parse this way, get the cannabinoids or the terpenes out of it. 503 01:10:14,480 --> 01:10:18,480 And as I was saying, we may have to adjust the parsing algorithm. 504 01:10:18,480 --> 01:10:22,480 Right. This one is just looking at the front page. 505 01:10:22,480 --> 01:10:28,480 But as we see here, the front page contains some of the information, 506 01:10:28,480 --> 01:10:37,480 but we'll have to go to some of the subsequent pages to actually get the terpene data. 507 01:10:37,480 --> 01:10:48,480 So once again, this parsing algorithm is going to need to be tailored so that maybe you say iterate over all of the pages 508 01:10:48,480 --> 01:10:53,480 until you've gotten all the data points that you're looking for. 509 01:10:53,480 --> 01:10:59,480 And once again, we're only getting, say, two or three terpenes here. 510 01:10:59,480 --> 01:11:04,480 Ideally, we could get all of the terpenes. 511 01:11:04,480 --> 01:11:11,480 So we're going to have to do this in a way that's clever prompting to do. 512 01:11:11,480 --> 01:11:17,480 So that's sort of the work that I've been tinkering on. 513 01:11:17,480 --> 01:11:32,480 You know, I encourage you all to play out, play around with this tool because I populated this with, as I said, north of 15,000 COAs. 514 01:11:32,480 --> 01:11:36,480 So that's why I'm kind of opening it up to you to kind of do development. 515 01:11:36,480 --> 01:11:48,480 I'm trying to sort of get an estimation about, you know, what is the marginal cost of making all this data searchable, parsable. 516 01:11:48,480 --> 01:11:58,480 And once it's parsed, then you'll have, you know, your nice data file here that awesome people like Rick and whoever. 517 01:11:58,480 --> 01:12:16,480 So you can now finally use Panabinoid Terpene data, all the data on the certificate, to build rich models and find amazing insights. 518 01:12:16,480 --> 01:12:21,480 So I encourage you all to play around with this. 519 01:12:21,480 --> 01:12:34,480 Submit issues, grab your keyboard and contribute, clone the code and use it in your own projects. 520 01:12:34,480 --> 01:12:36,480 Use it however you wish. 521 01:12:36,480 --> 01:12:42,480 And then if you scroll down here, you're also free to contribute. 522 01:12:42,480 --> 01:12:54,480 Because as I said, you know, I'm going to kind of be watching the usage here and I'm going to have to maybe think about, you know, 523 01:12:54,480 --> 01:13:05,480 if I make this tool widely available, it would be, I may unfortunately either have to charge for the service 524 01:13:05,480 --> 01:13:14,480 or if any of you want to contribute, then you're welcome to as well or run it on your own machine. 525 01:13:14,480 --> 01:13:17,480 And, you know, as I said, there's going to be a cost there. 526 01:13:17,480 --> 01:13:27,480 But as I said, there's a small marginal cost, but we've got the sources, we've got the methods, we've got the data. 527 01:13:27,480 --> 01:13:30,480 Let's dive into the analysis. 528 01:13:30,480 --> 01:13:46,480 So please feel free to contribute or explore or, you know, check out the GitHub and tinker to your heart's content. 529 01:13:46,480 --> 01:13:59,480 Or at the very least, give the repository a star if this is something that you find useful. 530 01:13:59,480 --> 01:14:02,480 What is everybody's thoughts, comments, questions? 531 01:14:02,480 --> 01:14:09,480 I know it was a lot of material, but it seems like we've finally achieved or not achieved, 532 01:14:09,480 --> 01:14:26,480 but we're finally starting to make progress on some tangible progress that people can put in their hands and use or put in their pipes and use. 533 01:14:26,480 --> 01:14:34,480 Did anybody have any thoughts they want to share? 534 01:14:34,480 --> 01:14:38,480 I needed that review, so thank you very much. 535 01:14:38,480 --> 01:14:42,480 It was good to, you know, for my first time back up for a long time. 536 01:14:42,480 --> 01:14:44,480 So thank you for that. 537 01:14:44,480 --> 01:14:50,480 But I'll go through the GitHub again, as well as the videos like we talked about. 538 01:14:50,480 --> 01:14:51,480 Thank you. 539 01:14:51,480 --> 01:15:00,480 And there's a lot there, and as I would really want to drive home, this is more of a demonstration. 540 01:15:00,480 --> 01:15:12,480 A lot of this needs to be perfected or if not perfected, probably, right, we can never get to 100 percent, but it can always be improved upon. 541 01:15:12,480 --> 01:15:24,480 So that was just a first iteration of what a COA parsing prompt could look like or a lab results search could look like. 542 01:15:24,480 --> 01:15:29,480 So these are all just a first iteration. 543 01:15:29,480 --> 01:15:38,480 And as I was pointing out, there's many, many imperfections and things that can be improved upon. 544 01:15:38,480 --> 01:15:53,480 So I encourage you all to use it, make it your own, tinker, explore, pester me if you want anything added to it or you want any data. 545 01:15:53,480 --> 01:15:57,480 So I think there's a lot there. 546 01:15:57,480 --> 01:16:01,480 We've just begun to explore. 547 01:16:01,480 --> 01:16:03,480 Rick, comment, question? 548 01:16:03,480 --> 01:16:21,480 No, I just wanted to say awesome presentation that I mean, you met or created a proof of concept, a really good one for a problem, a real problem, and use some cutting edge technology to do it. 549 01:16:21,480 --> 01:16:23,480 I think that you're right. 550 01:16:23,480 --> 01:16:30,480 You could improve the cost associated with it through like playing around with different models and stuff like that. 551 01:16:30,480 --> 01:16:44,480 I've been doing some like computer vision training and stuff, which you might that might be an application for this as well in terms of like collecting specific stuff and parsing it in a format that you want consistently. 552 01:16:44,480 --> 01:16:47,480 I'm going to take a look at the GitHub. 553 01:16:47,480 --> 01:16:52,480 I haven't seen it yet, so I'm excited and I'll try to contribute as much as I can. 554 01:16:52,480 --> 01:16:55,480 So thank you. 555 01:16:55,480 --> 01:17:03,480 Oh, phenomenal. And remember, this is actually only a cost we technically have to bear once. 556 01:17:03,480 --> 01:17:07,480 So once the COAs parsed, it's parsed. 557 01:17:07,480 --> 01:17:10,480 And so that's where you were mentioning your vector databases. 558 01:17:10,480 --> 01:17:13,480 And that's exactly what it can be used for. 559 01:17:13,480 --> 01:17:27,480 Parse the COA once that may cost one or two cents or say five to ten cents or what have you through OpenAI, but then save that in your vector database. 560 01:17:27,480 --> 01:17:34,480 So that way, if you see a PDF, remember we talked about hashes, it would be similar with a vector. 561 01:17:34,480 --> 01:17:43,480 So as soon as you see a PDF, you create a vector for it, you see if that vector already exists in your database. 562 01:17:43,480 --> 01:17:49,480 If it does, you can get the lab results without having to reparse it. 563 01:17:49,480 --> 01:17:56,480 If it doesn't yet exist, then you, you know, you parse it and save it in your database. 564 01:17:56,480 --> 01:17:59,480 So that could be a way you could save costs right off the bat. 565 01:17:59,480 --> 01:18:10,480 Well, if there's fifteen hundred of them out there right now, I'll play around with saving them to my vector database and then let you know how it works out. 566 01:18:10,480 --> 01:18:14,480 Is please I encourage you all to play play around with it. 567 01:18:14,480 --> 01:18:19,480 And as I was showing you, I just started with a sample size of five. 568 01:18:19,480 --> 01:18:26,480 And then as you saw, gradually worked up to a sample size of around fifty five. 569 01:18:26,480 --> 01:18:33,480 So as I was I always mentioned, I'm a small sample guy. 570 01:18:33,480 --> 01:18:35,480 You can do a lot with the small sample. 571 01:18:35,480 --> 01:18:41,480 So I encourage you to put five of these observations or put fifty in your vector database. 572 01:18:41,480 --> 01:18:48,480 If it proves useful, works well for your application, fill it up with fifteen hundred. 573 01:18:48,480 --> 01:18:52,480 But as I said, the cost may get a little high. 574 01:18:52,480 --> 01:19:05,480 But if there's value, I think hopefully there will be people out there who'd be willing to chip in a couple bucks to get their COA data. 575 01:19:05,480 --> 01:19:09,480 Phenomenal. Thank you all for coming. 576 01:19:09,480 --> 01:19:14,480 Your eyes, your ears, your attention. This is all moving the ball forward. 577 01:19:14,480 --> 01:19:21,480 This wouldn't happen without you. So you're the ones who are really helping this actually happen. 578 01:19:21,480 --> 01:19:34,480 And move everything forward. It's all the thoughts, the questions, the ideas that have happened throughout the whole cannabis data science meetup that's led us on this journey. 579 01:19:34,480 --> 01:19:42,480 You realize people wanted COAs. You realize there was a way we could parse them, get the text out of them. 580 01:19:42,480 --> 01:19:50,480 And we started exploring open AI tools and people wanted to learn more about those and learn more about terpene profiles and phenotypes. 581 01:19:50,480 --> 01:19:58,480 It's just all of this ties together until we are where we are today, where we realize, oh, there's Florida patients. 582 01:19:58,480 --> 01:20:04,480 They've been asking for their certificates like we encourage people to do. 583 01:20:04,480 --> 01:20:11,480 There's finally a lot of these URLs that are available because companies are making them available. 584 01:20:11,480 --> 01:20:22,480 That's my final point for today is we started asking and now we saw large companies like Cureleaf or putting these in VitaCann. 585 01:20:22,480 --> 01:20:27,480 They're putting them out by the hundreds, if not thousands. And it's becoming popular. 586 01:20:27,480 --> 01:20:35,480 So there's a company in Florida, The Flowery, and they do what's called COA drops. 587 01:20:35,480 --> 01:20:43,480 And so every few weeks or month, they'll just drop, say, 20 to 60 COAs. 588 01:20:43,480 --> 01:20:47,480 And that's their new inventory they're releasing at the store. 589 01:20:47,480 --> 01:20:55,480 And it's brilliant because we were talking about these tests can run up to, say, 500 bucks a pop. 590 01:20:55,480 --> 01:21:03,480 So if you tested 20 samples at 500 bucks a pop, that's like a $10,000 bill. 591 01:21:03,480 --> 01:21:11,480 And instead of just eating that bill, you actually get to use it for your benefit. 592 01:21:11,480 --> 01:21:18,480 So they're now using their COAs as a marketing tool. So they're like they're hyping it up. 593 01:21:18,480 --> 01:21:23,480 They're saying, OK, we're going to drop the COAs. And then they release them. 594 01:21:23,480 --> 01:21:29,480 And then everybody's excited about it. They're looking through them and they're talking about them on Reddit. 595 01:21:29,480 --> 01:21:36,480 They're saying, like, oh, like, check this one out. This one's got this one's got a ton of D-Lemonine or oh, watch out. 596 01:21:36,480 --> 01:21:49,480 This one's this one's a foreign bomb. So you can get quick feedback, get people excited for your products and use the lab result data to your advantage. 597 01:21:49,480 --> 01:21:57,480 Instead of just letting the COAs accumulate dust on somebody's old dusty hard drive. 598 01:21:57,480 --> 01:22:03,480 So I love it seeing companies putting this data to good use. 599 01:22:03,480 --> 01:22:10,480 It's getting consumers excited about it. And because of all of you awesome cannabis data scientists, 600 01:22:10,480 --> 01:22:19,480 we can add our piece to the puzzle and actually make the data readily accessible and analyzable. 601 01:22:19,480 --> 01:22:31,480 So and then it's up to all of you to start drawing nifty little insights. So what is the average for Nizine that you can expect? 602 01:22:31,480 --> 01:22:39,480 You know, what's the ranges on that? You know, what's the range on this beta pinin to D-Lemonine ratio? 603 01:22:39,480 --> 01:22:46,480 Like what's the strongest sativa in the Florida market? And what's the strongest indica? 604 01:22:46,480 --> 01:22:55,480 You know, you can start to discover these things and you know, you may find gold. 605 01:22:55,480 --> 01:23:02,480 Too cool. I'll get all these links and material posted and available for you. 606 01:23:02,480 --> 01:23:13,480 I encourage you all to to get in touch throughout the week, even if you want to use some of this material, you know, in your own projects or however you see fit. 607 01:23:13,480 --> 01:23:30,480 You know, let's just keep working on this, because as I said, as soon as I discovered this, all of a sudden, you know, now there's a, you know, a thousand and one more things to do that we can just keep slowly and methodically tinkering on it. 608 01:23:30,480 --> 01:23:36,480 Well, you need to go out and enjoy your day. I've kept you too long. 609 01:23:36,480 --> 01:23:44,480 Thank you one more time for coming, helping advance candidate science.