Description Transcript
Most data teams are sitting on a pile of unstructured GTM data — Gong calls, support tickets, product feedback — with no good way to surface it. The result: constant "can you pull this?" Slack pings that pull analysts away from real work.
Malcolm Angus, Analytics Engineer at Retool, solved this by building a self-serve Gong insights app that sales, CS, and marketing use everyday— no frontend skills, no borrowed engineering sprint.
We'll cover:
How to turn raw, unstructured call data into a searchable, summarizable app your GTM teams actually use.
How to apply the build pattern, from presets and filters to summarizing and trending, across any unstructured data source.
How shipping self-serve access to customer conversations drives faster GTM feedback loops and sharper coaching.
How to ship production-ready, securely governed apps without slowing down.
Whether you work in data or lead a data team, you'll leave with a pattern you can apply to your own stack.
Read more 0:01 All right, welcome everyone. We're 0:03 excited to have you here today and we're 0:05 going to be talking about how data teams 0:08 have built production ready customer 0:10 insight applications in Retool. 0:12 So we're going to do things a little bit 0:14 different here today, but 0:16 no heavy slides and instead in the next 0:18 30 minutes we're going to be spending a 0:20 lot of time on the production ready 0:22 customer insights app that was built by 0:25 a data team 0:26 and it's really built for a gold market 0:29 pipeline that's being used in Retool 0:31 today. 0:32 Hopefully you'll also walk away with a 0:33 blueprint to also replicate this. So 0:36 this is an excellent example of how we 0:38 dog food here at Retool and it's an 0:41 internal app that our teams actually use 0:43 on a weekly basis. But just some quick 0:45 housekeeping before we dive in. A 0:49 reminder is we are recording this 0:51 session and there will be an email out 0:54 after the 0:55 of the recording itself and if you do 0:58 have any questions during the 1:00 conversation today, you can submit them 1:02 through the Q&A feature. So you should 1:04 see that down at the bottom of the zoom 1:07 screen. 1:08 And we will try to get to those 1:11 questions 1:12 at the end. But hi, I I am Elise. I'm a 1:16 solutions engineer at Retool and joining 1:19 me today is Malcolm and he's the one who 1:21 actually built what you're going to 1:23 you're going to be seeing here in a 1:24 little bit. So Malcolm, say hello. Yeah, 1:26 hi everyone. My name is Malcolm. I'm an 1:28 analytics engineer at Retool. 1:31 Fun fact is this is actually my second 1:33 webinar with Retool. I did one several 1:35 years ago, but actually as a champion 1:37 customer. I'm super excited to show you 1:39 guys what I built. 1:42 Okay, so let's get started. We have a 1:45 tight timeline today. 1:47 So here's how we'll be submitting our 1:49 time together 1:50 together today. We'll start with a quick 1:53 honest look at the problems that are 1:55 affecting the teams across the board 1:57 right now. So not just data teams. But 2:00 then Malcolm's going to walk through the 2:02 actual application and the architecture 2:04 behind it. So we'll close with outcomes 2:07 and Q&A. So let's get started. 2:11 And now before we actually dive into the 2:13 solution itself, I do want to spend a 2:16 quick 60 seconds on something and and 2:18 it's you know, the odds are that at 2:21 least one of these will sound familiar 2:23 to someone in this room and I want to be 2:25 clear that this isn't just a data team 2:27 problem, but we are hearing versions of 2:29 this from marketing and product RevOps 2:32 and the C-suite. So the first is what we 2:35 call this dashboard factory. 2:38 The dashboard factory is essentially a 2:41 capacity trap. Right? Every stakeholder 2:43 request for a new report or a tweak ends 2:46 up pulling the most skilled people away 2:48 from building the things that actually 2:49 move the business. 2:51 So the team is busy, but they're not 2:53 actually building reports and 2:55 applications for the long term. So for 2:58 leaders in the room, this is the reason 3:00 that the data teams always seem like 3:02 they're underwater, but they still can't 3:04 get a straight answer fast enough. 3:07 The second one is this read-only 3:09 insights. So every time someone finds 3:12 something important, turning it into an 3:15 action means filing a ticket or waiting 3:19 a few days after that and then that 3:21 moment passes and ultimately the 3:23 momentum dies and the loop never closes. 3:26 So for RevOps and go-to-market leaders, 3:29 this is the gap between knowing what's 3:31 happening and being able to actually do 3:34 something about it and take those 3:35 actions. 3:37 The third is manual processes. So 3:40 companies have made these massive 3:42 investments into 3:44 modern infrastructure like 3:47 Databricks or Snowflake, 3:49 these data warehouses among others, 3:52 but on the ground people are still using 3:54 and living in Excel because the tools 3:57 don't connect to the actual workflows. 3:59 And that's not just the efficiency 4:01 problem that we're seeing, but it it 4:03 also becomes this governance nightmare. 4:05 So 4:06 what makes this especially painful is 4:09 what happens when tries teams try to fix 4:11 it themselves. 4:13 They might be reaching for whatever is 4:15 available, right? Maybe that's front-end 4:17 or back-end engineers or turning to 4:20 traditional BI tools like Power BI, 4:23 Tableau and trying to actually support 4:27 write backs, but this isn't the reality 4:30 today and even doing things like 4:32 open-ended AI analysis 4:34 is something that you know, we're all 4:35 moving towards and striving towards 4:37 today. 4:39 Also they might end up trying to look at 4:42 coding platforms like Lovable, 4:44 you know, do some vibe coding to produce 4:47 some beautiful applications and they 4:49 look production ready, but falls apart 4:53 the moment you try to do some real 4:54 functionality or add in security or 4:57 governance. So that last mile that you 4:59 need to actually get production grade 5:01 applications is not actually getting 5:04 people where they need to. 5:05 So everyone right now is trying to solve 5:08 this themselves and they keep running 5:10 into some some walls, which is what 5:13 Malcolm is going to address here next. 5:16 Thanks Elise. So yeah, you know, I 5:19 wasn't actually the first person to try 5:20 and build this Gong app internally and 5:22 I've talked to some other companies who 5:24 have been trying to build more or less 5:26 the same app cuz as some of you probably 5:28 know like Gong has a great product, but 5:31 some of the UI you wish you could do 5:32 kind of that, you know, last mile 5:34 analytics across many different calls. 5:36 So like I said, I'm not the first person 5:38 to try and solve this, but I noticed 5:40 basically there was a graveyard of apps 5:43 that never really made it to production 5:45 for a few reasons. And I you know, this 5:47 title is at first was called why, you 5:50 know, where data teams get stuck 5:51 building data apps, but in 2026, I think 5:54 it's more real to say, you know, where 5:56 teams get stuck building data 5:58 applications, right? Cuz almost anybody 6:00 can build these kinds of apps, 6:02 especially with the new technologies 6:04 available to us. But I keep saying, you 6:06 know, seeing the same problem as a data 6:08 professional and so I want to help 6:10 unlock you and think about like, okay, 6:12 how does a data professional think about 6:14 some of the tradeoffs and how to get 6:16 these apps into production. So the first 6:18 kind of trap people fall into is not 6:21 understanding understanding data 6:23 proximity and data strategy. There's a 6:25 statistic like only 92% of companies 6:28 that invest in AI, only one of them 6:30 actually kind of get apps to production. 6:32 A huge part of this is because data's 6:34 just scattered everywhere and kind of 6:36 half-baked AI applications. So either 6:39 you're really familiar with the 6:41 underlying shape of the data that's 6:42 coming from say these different data 6:45 sources like Gong's API and you have 6:48 some idea of how to build data products. 6:50 You just can't do it halfway and expect 6:53 a a fully functioning application. 6:56 The second trap I kind of see, 6:58 especially non-data professionals run 7:00 into, is they're trying to do too much 7:02 in the application or the app layer 7:04 itself. So for example, in Gong, right? 7:07 We have to make API calls to fetch the 7:10 data. We have to join the data from the 7:12 various different Gong endpoints. Then 7:15 we might want to do some joining into 7:16 Salesforce data and then finally do the 7:19 enrichment. This is just too much 7:22 regardless of any platform that you've 7:24 selected for the application layer 7:26 alone. I'll get into how we kind of 7:28 solve for this. 7:29 The next one is traditional BI. So like 7:32 Elise mentioned, if you're trying to 7:33 build these these data applications, 7:36 especially AI-powered ones in, you know, 7:38 traditional BI like Tableau or Power BI, 7:41 oftentimes they're just not supporting 7:43 these last mile open-ended LLM analysis. 7:46 I haven't seen them make, you know, be 7:48 able to make API calls to say OpenAI's 7:52 LLM's API or Gemini or whatever, right? 7:56 The other point I want to make here is 7:58 traditional BI doesn't have as rich of a 8:00 UI component library unlike Retool, 8:03 where you can't really own every single 8:05 pixel and you can't own the user 8:07 experience. So often times it leads to 8:09 more cluttered dashboards than what 8:11 feels like a real application. 8:14 And the elephant in the room, the last 8:15 one is, okay, why not just vibe code 8:18 these? You certainly can, but if you I 8:20 I've seen this in other companies where 8:22 they try and build it in Replit and 8:24 Lovable, they get pretty stuck where 8:26 they they wind up with a very beautiful, 8:28 you know, UI that is more or less looks 8:31 like what it can do to do the Gong 8:33 analysis, but really it lacks actual 8:36 functionality and these apps never make 8:37 it into production because often times 8:40 you have to get it deployed in in your 8:42 cloud and on your data and that just 8:44 leads to endless kind of engineering 8:45 ticket queues back and forth. 8:49 Okay. 8:50 Yeah, and I think that's a good segue 8:52 into, 8:53 you know, how do we do all this in 8:55 Retool? So I wanted to just give a full 8:56 picture as to what Retool does offer for 9:00 anybody that is new to Retool. 9:02 So at a high level Retool is an 9:04 enterprise app gen platform. So it lets 9:07 any team like data, RevOps, product to 9:11 ship these working apps and automations 9:13 in hours and not weeks. And there are 9:16 three core building blocks. So the first 9:19 being apps for web and mobile and this 9:21 is how teams build dashboards, 9:23 operational tools, portals or internal 9:26 products that people actually use, 9:29 which will be like the the real 9:31 interactive applications. 9:33 Second [clears throat] 9:34 is this automated workflows and this is 9:36 where you connect systems. You can run 9:38 logic and handle approvals 9:41 and automate an end-to-end process. 9:44 And then the third is the AI-powered 9:46 agents and this is how you bring AI into 9:50 operational workflows in a governed way. 9:52 So rather than this standalone 9:54 experiment, 9:56 you know, you have something that you 9:57 can actually put into production. 9:59 Uh, but where Retool is particularly 10:02 powerful is that you're not locked into 10:05 a drag-and-drop box. Uh, Retool lets you 10:08 write JavaScript or SQL, Python, and 10:11 those three languages that, uh, you 10:13 know, teams already typically know. So, 10:16 uh, you're able to write that directly 10:18 within the Retool platform. 10:20 Uh, 10:21 ultimately, you end up getting the speed 10:23 of this visual builder with the full 10:25 flexibility of the code when you need 10:27 it. So, there's no new language you have 10:30 to learn, and there's less context 10:32 switching, and you don't have to be that 10:34 front-end developer, and you can even 10:36 leverage, uh, natural language. So, you 10:39 can just describe what you want, you can 10:41 build or change a particular element in 10:43 the Retool apps, uh, and this this will 10:45 do, um, you know, everything from, uh, 10:48 generating SQL queries and JavaScript 10:51 into, uh, generating entire UIs. 10:55 Uh, so, with that all in mind, um, 10:58 I I do want to mention that Retool is 11:00 also completely data agnostic. So, 11:01 whether that data lives in Snowflake or 11:03 Databricks, uh, or REST API, uh, Retool 11:07 connects to those using pre-built 11:09 integrations out of the box. Uh, same 11:11 goes for AI, you're not locked into one 11:13 model. You can connect to OpenAI, 11:16 Anthropic, Google, or any other LLM your 11:19 team is already using, uh, or wants to 11:21 use. Um, and then lastly, you can 11:23 integrate with your SDLC through source 11:25 control. And so, every app ships ships 11:28 with, uh, SSO, audit trails, role-based 11:32 based access control, secrets 11:34 management, uh, from day one. And that 11:36 security and governance, uh, isn't going 11:38 to be, um, an afterthought, and it's a 11:41 foundational element. So, for anybody 11:43 that's managing data across teams or 11:46 worrying about who has access to what, 11:48 uh, that will that will matter a lot. 11:51 Uh, but now that you understand the 11:52 underlying platform, Malcolm is actually 11:55 going to take us through the 11:56 architecture of the app and the demo of 11:58 what he built. 12:03 Thanks, Elise. Okay, so before I get 12:05 into like the nitty-gritty details of 12:07 the app, again, I have to mention that 12:09 the reason a lot of these apps didn't 12:12 make it to production is cuz they didn't 12:14 focus on the underlying data pipeline, 12:17 and instead did too much, uh, you know, 12:20 data processing in the application 12:21 layer. So, here's how I kind of decouple 12:23 the two, right? So, a lot of you are 12:25 probably familiar, like, again, with 12:27 this Gong problem, or somewhat familiar 12:29 with the Gong API. So, I'm going to tell 12:32 you basically how we I took that data 12:34 and then preprocessed it to make a very 12:36 friendly kind of data product, or rather 12:38 tables, so the Retool app can can query 12:41 easily. So, you might be familiar, 12:43 there's a few different endpoints, like 12:45 getting the list of calls, getting 12:47 specific metadata for a given call, and 12:49 then getting the transcripts for a given 12:51 call. Uh, so basically, I have a process 12:53 where, you know, I'm an app engineer, 12:55 this is normal, you know, part of the 12:57 job or for a data engineer. We're 12:59 syncing that data into a data warehouse, 13:01 in this case, we use Databricks. So, we 13:03 wind up with two tables, call metadata 13:06 and then call transcripts, and then we 13:08 have other data sources like Salesforce, 13:11 where we want to basically join that 13:13 data together to have even more enriched 13:15 data that is not just available in Gong, 13:18 right? So, the the kind of final data 13:20 product here is 13:22 two tables, or rather one table, which 13:24 is these enriched calls that include all 13:27 the metadata from Gong, all the metadata 13:29 from Salesforce, and then also this this 13:32 vector index, which is, uh, 13:35 if you're familiar with vectors, is is a 13:37 way to enable rag, right? We're 13:38 basically doing semantic search. So, 13:41 here we can do semantic search against 13:43 those transcripts pretty easily. The 13:45 last last step here is basically, okay, 13:47 now Retool app doesn't have to worry 13:49 about any of this process, it can query 13:51 these directly. 13:52 I'm going to the next slide. So, if we 13:54 double-click into Okay, now the app 13:56 itself, how does the app function, 13:58 right? So, as I mentioned the previous 14:00 slide, we have these enriched calls, we 14:02 have the vector index, that's in 14:04 Databricks. Um, 14:06 now, the application itself is basically 14:09 allowing the user to filter for various 14:12 types of calls. I'll show you what that 14:14 actually looks like in practice, but 14:16 that's basically a search hitting these, 14:19 uh, 14:19 these, uh, this calls table, right? And 14:22 then we get back the the match calls. 14:24 Then this kind of magic happens where 14:27 the, pardon me, so the transcripts, 14:30 again, are in these enriched calls, and 14:32 we basically forward those via Retool AI 14:36 as the last-mile LLM analysis, okay? The 14:39 magic about Retool AI is it's, again, 14:42 like LLM agnostic, and basically, you 14:45 can pick whatever LLM provider and model 14:48 that you want, uh, for whatever purpose 14:50 you need, okay? So, for example, there's 14:53 actually a drop-down I'm going to show 14:54 you. We've defaulted to use OpenAI's LLM 14:57 in this case, in this application, but 14:59 you could easily switch that out, and 15:01 you can also switch the model. So, the 15:03 actual end user of the application can 15:05 decide, okay, uh, for this specific use 15:08 case in question that I'm asking, I 15:09 actually want to use this this model 15:11 instead of this other one. So, 15:14 Retool, again, is basically acting like 15:16 Retool AI is acting as this broker, and 15:17 will return that analysis, and then 15:19 finally, we'll pass it to the front end 15:21 and render it. So, that's a little bit 15:23 how it works, um, in terms of the 15:25 application, and I'm excited to actually 15:27 show it to you. 15:29 >> [snorts] 15:29 >> All right, I'm going to take over here. 15:35 All right. So, uh, before you guys get 15:38 too excited and read too much into the 15:40 data, just you know, this is a demo 15:41 application with with dummy data. So, I 15:44 wasn't able to fully replicate the 15:46 entire experience for you, uh, but 15:48 you're going to have to use your 15:49 imagination a little bit, okay? So, the 15:51 Gong super app, um, keep in mind that 15:54 this is an application that is designed 15:56 for many different kind of internal 15:58 personas, whether it's product, 16:00 marketing, sales, so I had to keep 16:02 somewhat of a a broad, uh, you know, 16:05 user experience. Um, so one thing in 16:07 particular that right away is helpful 16:09 for users is is I know it looks simple, 16:12 but having this watch tutorial video 16:14 right in the application itself is, 16:17 like, saving a bunch of back and forth, 16:19 um, with me trying to explain stuff and 16:21 having to link and support different 16:23 Google Docs or, you know, share access 16:25 and whatnot. Um, okay. So, you'll notice 16:28 that there's the kind of introductory 16:30 part of the the experience here is the 16:33 search capabilities. So, we're able to 16:35 filter calls, right, using this, 16:38 uh, metadata based on the call date, so 16:41 start and end date, right? And then we 16:43 have 16:44 transcript regex. So, if you guys are 16:46 familiar with regex or regular 16:48 expressions, it basically allows you to 16:50 do, uh, various kinds of of text 16:52 matching. So, in this case, it's it's 16:54 somewhat of a exact match search, right? 16:57 And a lot of people, even myself, 16:59 aren't, you know, regex masters, so I 17:01 have this regex help here, so you can 17:03 click on that, and then in plain 17:05 English, you can describe, uh, what are 17:07 you trying to match, right? I'm trying 17:08 to match React and some of these other 17:10 keywords, and you can ask AI, and under 17:13 the hood, Retool AI will basically, 17:16 uh, you know, fetch from the LLM, hey, 17:18 I'm trying to do this with regex, can 17:20 you please translate it into some regex 17:22 pattern? You can copy and paste it in in 17:24 put it here. If that's not enough, we 17:26 also have like the option for users to 17:28 select kind of what I decided is 17:30 pre-populated AI regex, like, hey, we're 17:32 trying to match transcripts where people 17:34 are talking about AI, and here's the 17:36 different keywords for AI. Or, for 17:38 example, let's match the regex for 17:40 different competitors, right? So, that 17:42 will pre-populate, uh, this field here. 17:45 Now, you might, again, think back to the 17:47 vector index that we set up to do kind 17:49 of some rag analysis. Again, we want to 17:52 say, okay, uh, 17:54 some customers might not saying AI 17:56 explicitly, but kind of talk about it 17:58 indirectly, right? They use different 18:00 language to, you know, mention AI. So, 18:02 this is where the semantic search plays 18:04 a role, right? So, if you search for AI 18:06 here, it'll return, uh, call transcripts 18:10 that are like AI. So, that's super 18:12 helpful. 18:13 Last thing here is job title regex. 18:15 Again, we have enriched data from 18:17 Salesforce on the opportunities, on the 18:20 contacts, and then looking at the 18:22 contacts, we know, hey, this is that 18:23 person's job title, so this is where we 18:26 can do that filtering as well. Uh, 18:28 you'll notice also here there's a toggle 18:30 for external only. So, a big kind of 18:32 feature request we got from the users 18:35 is, okay, I actually want to match what 18:37 customers are saying, not necessarily 18:39 what Retool employees or sales, uh, are 18:42 saying, right? So, we want to be able to 18:44 match in the customer's own language, 18:46 um, and so that provides this kind of 18:48 functionality. So, for example, we can 18:50 search those calls, 18:52 we can see the, uh, call results here, 18:55 right? Again, this is, you know, demo 18:57 data, but you can see all of the 18:59 different kind of blended blended 19:01 sources, um, with with Gong and with 19:04 Salesforce. Um, one nice thing, this is 19:07 a a native Retool, you know, table. You 19:09 can easily filter it across all the 19:12 different columns and sort it, and 19:14 there's a lot of functionality there. 19:16 So, uh, we select one of these calls, we 19:19 can open the call details, and this is 19:21 where it gets into like feeling like a 19:22 real application and not just kind of, 19:24 you know, a crowded dashboard. You can 19:26 see all the metadata for the call, you 19:28 can see the transcript, all these links 19:30 automatically deep link to Gong or to 19:33 Salesforce. Um, again, [snorts] I wish I 19:35 could show you like some of the more 19:37 functionality here with the transcripts, 19:39 but unfortunately, they just isn't 19:41 possible with this this demo data. 19:44 Okay, so the real meat and the potatoes 19:46 of this dashboard is actually the 19:47 ability to select multiple calls, right? 19:49 Let's say you had a hypothesis about uh 19:52 certain segment of customers, specific 19:54 vertical, right? Uh we can select up to 19:58 I think I got it up to like 40 calls at 20:00 once depending on the the specific 20:02 model. But let's say for example we 20:04 select these eight calls, you scroll 20:06 down here and it says summarize eight 20:09 calls. You know, notice that it's 20:11 virtually in real time uh that we can 20:13 get this AI summary. And this is because 20:16 I'm actually using um one of the the 20:18 various models here. So you can see I'm 20:21 giving the end user the ability to 20:23 select, okay, within Open AI what model 20:26 do you want to uh and analyze these 20:28 eight calls for? Okay? So that's a brief 20:32 of of how the summary works, but you 20:34 know, that's summary's great, but it's 20:36 not enough for most end users. So we 20:39 have this kind of extra tab here that 20:41 allows for this custom prompt. So the 20:43 idea here is okay, you can ask any 20:45 question that you possibly want. In this 20:47 case like let's maybe I'm a marketer and 20:49 I want to understand okay, what are the 20:50 or product and I want to understand the 20:52 biggest pain points that our customers 20:54 are facing. So we can ask the AI and 20:57 then get back in real time a curated 20:59 response for those calls with respect to 21:03 the specific prompt. 21:05 Cool. So that's kind of the magic of 21:06 this and it comes down to what you know, 21:09 what is the inspiration and and 21:11 creativity for the end user in terms of 21:14 the value they can extract. But 21:15 hopefully you can already see kind of 21:16 like how this creates almost like a 21:18 research function out of the the data 21:21 and is really expanding what data teams 21:23 can do moving from the quantitative to 21:26 the qualitative the value add, right? 21:29 But we still want to address, you know, 21:30 some some quantitative here. So for 21:32 example, uh imagine that we have 21:35 multiple, you know, transcript uh regex 21:38 uh keywords that we want to match for. 21:40 In this case it would render split by 21:43 each keyword of the volume of calls that 21:46 are mentioning those those keywords, 21:47 okay? So let's say for example we want 21:49 to track um 21:51 competitor mentions, right? Across each 21:53 competitor, uh how many times in both 21:56 volume, but also the uh distribution of 21:59 mentions across each competitor. So we 22:01 can do that here. We can see the 22:02 percentage of calls, again, you have to 22:04 use your imagination a little bit for 22:05 this demo data. 22:06 Um and then let's say we want to 22:08 increase the kind of granularity, we can 22:10 look at okay, what are the you know, 22:12 number of opportunities that are 22:13 mentioning competitors, right? So that's 22:16 the gist of of basically what the user 22:18 experience is like and I'm going to jump 22:20 in real quick to 22:21 uh like what's going on under the hood 22:24 um 22:25 of how this app is is actually kind of 22:27 wired up, right? So a few things to note 22:30 depending on your level of familiarity 22:32 with uh with Retool, um all of these are 22:35 basically various UI components. So 22:38 Retool provides you a very rich library 22:40 of UI components 22:42 um and like I actually created this app 22:46 a while ago and this is before we had 22:48 our Assist product, which basically 22:50 allows you to do uh text um almost vibe 22:53 coding experience. So you can basically 22:55 edit this app now using pure text, you 22:58 don't have to know much about the app or 23:01 or Retool, right? And you can also ask 23:03 questions against uh against the app as 23:06 well. Similar experience if you want to 23:08 build the app from scratch, you totally 23:09 could too through this interface. But 23:12 let's go back here, you can see that 23:13 this is a you know, a date range, right? 23:15 All these are different input text 23:16 fields. Um this table for example, I'll 23:19 show you how it it's kind of wired up. 23:21 So if we go into the inspector, we can 23:23 see okay, this table is being populated 23:24 by the search and if we open that, we 23:27 can see oh, this is actually a query and 23:30 it's somewhat of a complex query, right? 23:32 I mean, I write SQL all day, it's no big 23:33 problem, but someone else might not be 23:35 as familiar and that's fine too. So we 23:37 have this ask AI where you can actually 23:39 edit, you know, using AI to edit the 23:41 query or write a new query altogether. 23:44 But you'll see within this query, right? 23:46 We can see the date range, you know, 23:48 start and end dates. This is how it's is 23:50 referencing the various aspects of of 23:52 UI. 23:53 Uh we can see the job title regex and 23:55 more. 23:56 So this is a a gist of how it works, 23:58 right? 23:59 Um and I'm going to show you for example 24:02 that this search is actually part of a 24:04 few other uh queries to Databricks, 24:07 right? So this is our underlying table 24:10 that I I talked a lot about is like this 24:12 enrich table that is where all the data 24:15 lives. We have a few other queries that 24:17 are powering this experience, but then 24:20 we'll see if we scroll down to that 24:21 summary. Let me hide this inspector, we 24:23 can see the summary here. Open up the 24:25 inspector again. 24:27 We'll see that this summary is actually 24:29 being powered by 24:31 this uh Retool AI uh uh resource. So 24:35 this is allowing us to basically 24:37 generate text with the given input, 24:39 which is the transcripts for the LLM. 24:42 It's doing a little bit of like 24:43 JavaScript, again, the Retool AI has 24:45 helped write that JavaScript cuz you 24:46 have to do that last mile formatting uh 24:49 for the LLM to actually like being able 24:51 to to read it easily. So then you can 24:54 also select the model here. In this case 24:56 we've hardcoded Open AI, but I've 24:58 allowed the user to input and uh 25:01 basically reference the the actual UI 25:04 where the uh model is selected and then 25:06 I have a system message. I just came up 25:08 with this. So for example, you know, 25:10 summarize [clears throat] the following 25:11 calls between Retool and customers. If 25:13 there's multiple calls, give a high 25:14 level aggregate summary. Last step here 25:16 is this temperature zero, just saying 25:17 hey, 25:18 no creativity. We want more or less 25:20 predictive uh inputs and outputs, right? 25:23 So I'm going to pause there 25:25 um and I will go back to oops, sorry, 25:28 I'm actually going to share my screen 25:30 again. 25:31 Um 25:31 Er at least do you mind taking over in 25:33 the in the slides? 25:35 Yep, for sure thing. I will share my 25:37 screen. 25:38 Um yeah, thanks Malcolm for walking us 25:40 through that. Um 25:42 you know, we were trying to answer some 25:43 questions in the chat live, but um we 25:46 will try to hold them here for the end. 25:48 But uh with that uh Malcolm, I'll I'll 25:50 hand it back to you. Yeah. So I'll keep 25:53 this brief. This is basically the you 25:55 know, life after this app was 25:56 introduced. It it really reduced the 25:59 number of like ad hoc requests from our 26:01 team to be able to pull Gong data. 26:03 Really enabled, you know, self-service 26:05 for all the different uh business 26:07 stakeholders. And then this is a kind of 26:09 uh proof of concept where hey, we can 26:11 launch other kinds of qualitative 26:14 uh applications as a data team and 26:16 continuously, you know, have a progress 26:19 forward um in terms of of what we can 26:22 ship, but also inspired other, right, 26:24 non-technical 26:26 uh stakeholders to build their own apps 26:28 as well. Again, like our team we should 26:30 be focusing as as data team on building 26:32 the data products and then the actual 26:34 applications can be built by other 26:36 teams. Um so thus reducing the 26:39 bottleneck on us. 26:42 Okay, I think we can have some time for 26:45 Q&A. 26:48 Okay, um 26:49 so let's get to some of those questions. 26:52 Um so let me see here, we have a few uh 26:56 queued up here. 26:57 Um 26:58 yeah, I think some people have been 27:00 using the Q&A feature, others have been 27:02 putting it in chat. So we're going to 27:04 look through the Q&A uh first. 27:07 Um so one of the questions we have is 27:10 how do you handle data governance when 27:12 you're giving non-technical users direct 27:14 access to sensitive uh call data? Uh 27:18 yeah, that's a great question. Um you 27:20 know, as I mentioned before, that 27:21 governance piece is definitely a key 27:23 part into making sure that you're having 27:25 production grade enterprise applications 27:28 and there's a few ways that that can be 27:29 approached. Um often times this can be 27:32 at the data level. So when querying for 27:35 uh data sources, this can be something 27:36 that um you know, on the data level or 27:39 on that resource level, we can respect 27:42 the permissions of the authentication of 27:44 that third party. So for example, if you 27:46 are using something like Databricks, 27:48 they have permissions uh and OAuth that 27:50 we can essentially follow. Uh but built 27:53 into Retool are mechanisms like source 27:55 control and role-based access controls 27:58 that you can also make sure that the 28:00 apps and the resources you connect to um 28:04 are also um 28:05 you know, um 28:07 uh a granularly controlled to that 28:09 specific individual or group. 28:12 Uh so good question. 28:14 Um another one is how long did this take 28:17 to build? Um so Malcolm, I'll I'll have 28:20 you take that one. 28:22 Yeah, uh it's a great question. I think 28:25 the application itself was very easy to 28:28 create initial prototype and prove its 28:30 value, but the real time that you have 28:34 to invest is in that data pipeline part. 28:36 Like you have to understand okay, what 28:38 is the data and various endpoints we're 28:40 getting from Gong, the shape of the 28:42 data, how do you join it together, how 28:44 do you join it to Salesforce 28:46 um 28:47 and and uh how do you create the vector 28:48 indexes? Like that's that's honestly the 28:51 the the struggle. Um it's not in the 28:53 application building. You can build the 28:54 app itself if you got the data right and 28:57 like a few hours. And with uh AI, you 29:00 know, uh Retool AI now um or I should 29:04 say Assist, like more like less than an 29:06 hour. 29:07 Um so it it really depends on on the 29:10 state of your data. 29:14 Okay, and then we do have um another 29:16 question. I see this repeating, but um 29:18 can we get an export of this app? I'd 29:20 really like to review and possibly 29:22 replicate. 29:24 Um I am probably not the right person to 29:28 figure that out, but I can get the 29:30 answer. 29:31 Yep, we'll we'll see if that's something 29:33 we can uh templatize and give out to 29:34 people. 29:36 Um but uh we are kind of at time here. 29:39 So I I will uh make sure that any other 29:42 questions we did not address today, 29:45 we'll we'll follow up via email. 29:47 Um but also to kind of to to wrap up and 29:50 end, there are just a few things. So 29:53 for for anybody who does have questions 29:56 or you want to kind of see and 29:57 understand how this would interact with 30:00 your stack and your data sources. You 30:02 know, there's a link to a call that 30:04 we'll send that you can set up. 30:06 And we'll also you know, talk through a 30:09 great starting point 30:11 if if that's something you're interested 30:12 in. 30:13 But also if there's 30:16 you know, there's resonance in what you 30:18 saw today and you're interested in more 30:21 so taking these insights and taking it 30:24 to the next level. We do have a great 30:26 webinar 30:27 that our manager of developer relations 30:30 walked through specifically with 30:33 you know, automations and workflows 30:36 and how you can turn those insights into 30:37 action. So I would highly recommend 30:39 actually watching this recorded webinar 30:43 that we had back in January. So you can 30:46 scan that QR code on your screen or you 30:49 know, we'll we'll drop that link I think 30:51 in in the chat as well. So 30:54 really appreciate you being here today 30:55 and thanks for your time and let's go 30:58 build great software. 31:00 Thank you everyone.
