afternoon-opus

# Transcript: 2026-05-05

> 1 time blocks from 12:31 PM to 1:18 PM

---

### Casual work chat with colleague
**12:31 PM - 1:18 PM PDT** | *meeting*

**Microphone:**
Both of us zipping around doing shit. I hadn't seen David frantically messaging us about something going on, so hopefully that's good.

Why would you fucking say that, man? You don't say that because now you're inviting the universe to give us that shit.

Ah, damn it. I jinxed it. Yeah, I saw some H27s but they're actually on the fucking V2 dashboard, so the quicker we can get rid of the old V2s, the better.

Yeah, I can only imagine that happening. We'll get rid of it soon.

And then I'm sorry about the invite bullshit again. To get us started on this relations or this agent stuff the next day or two, and then I'm just gonna tell David we need to take a week. I put together a couple tickets for completely removing all the IAM stuff from SuperConnector and just rolling out complete WorkOS integration, invitations and all.

Nice. That would be great. If we could just do that, that stops one of the largest internal pain points for him.

Yeah, I think that would do wonders for us.

Everyone's talking about the game. At the point where I'm now literally pasting all my emails into Claude, versus being like, am I coming across weird here?

Yeah. I almost always did that for the last three years probably because I've gotten so many comments about how everything I wrote was so verbose, which is really funny because I had to go to AI, which is like the most verbose tool I could possibly think of, to find ways to reduce my words.

The whole Claude m-dash thing. I have been using m-dashes since high school. I really care about m-dashes. The other thing is, you know, the proper use of m-dashes. And so it's hard because I now have to take them out.

I can't remember what exactly it was, but I was like, I have to know why. Why do these LLMs need to use the m-dash like this? And so I asked ChatGPT because it was obviously the worst of the bunch. And it goes on to say that it's like a verbal "um" for people, like a filler word basically, that allows it to sort of say something but not mean it fully, I guess. And I'm not sure if it's a product of the way that they curate their training data or what, but I guess it's sort of like, here's a sentence fragment that you can sort of take or leave, as well as this weird sort of rhythmic beat that the LLM hits when it's outputting tokens.

It's interesting because I don't mind the m-dash. I mind it because I don't want to change my emails. I don't like having to change my style of typing or writing. But I don't mind it. I'd be really pissed if it started replacing m-dashes with comma space "um" comma, you know, which it obviously would never do, but that's what you just said. It's the same thing, right? I'd be super pissed. I'm like, you're clearly wasting tokens. You don't need to do that.

Has Claude started swearing anymore? Because mine keeps going, "shit, I messed that up."

A lot of people have been reporting that 4.7 is swearing a bunch.

Interesting. Maybe. I'm still incredibly happy with Claude for what it's worth. Like, you know, Todd's comment this morning at standup — it's like a small thing to get upset about, and yet. Like, sure, but I'm not not buying it.

Well, Todd isn't really using Claude, to my understanding. He seems to be using GLM 4.7 with a little hint of Claude.

Oh. But they're all distilled on Claude anyway, so it's not like it's all that different. But yeah, no, I mean I could have told him that was a bad idea, but...

You probably can't tell him that actually, but it's cute that you think so.

Yeah. Okay. So I have to be careful that I don't make jokes like this too much or people are gonna think I'm not good at my job. But doing what I normally do — not my job — and just having Claude do it, I took our conversation and post and had Claude read it plus my handover docs I had generated from the prototype development.

And so what I really had to do, and I'll share this with you though — I don't even know that we want to keep this in the document. I think we want to solidify this and then make each one of these into a Linear ticket in a project. But what I had to do was just break down all of the things we had talked about essentially, and then the rest of the documents start kind of trying to figure out what it looks like.

And so that map account network is really good given how fast the connector dashboard works, especially when Sidekick is actually running. Elasticsearch is the way. Yeah, Elasticsearch I think I'm starting to really feel is the way.

I don't know. So I think you just want to go through these and just kind of talk it through.

Mm-hmm. I've got mine as well.

Okay, I'm gonna... Yeah, absolutely. I also did something like a similar exercise, but okay, I'm happy to just work through with you. I'll message you the link. You can pop it on your screen if you want. Just wrapping up some changes. Everything was referring to you as a collaborator instead of Michael.

Yeah, so okay. Map account network is kind of continuous, you know. It takes the idea that the target company ID is an odd way to describe the list.

Funny feeling, yeah. But sometimes Granola doesn't — I must mumble or something. I mean, I found that transcription can only go so far, but the quality of your mic is going to be the deciding factor in a lot of this.

Yeah, I thought a lot of it. Okay, so the output — oh, I'm on the normal one, that's right. Thanks. So the target ID, account ID, and returns relationship IDs. So this is just continuous rule rerunning on a cadence, you know, and on relationship graph deltas. That sounds right. Doesn't need context or can be fast.

Writes to Rails for future reference. I'm not sure how much I'm going to point to it.

Okay. Yeah, and so this is coming from the process probably or something of that nature. Makes sense, right?

Yep.

Now, key for Snell — this has got it wrong, I think.

I think this has to come from Lima data — something like prospect company or prospect employees. You could start with seniorities and titles, which I think makes sense. We can also potentially use a combination of both the existing data that we have plus Lima data and map prospects that fit the buying target criteria for the account.

Cool, I agree with that. And I think find board members and advisors — so we could do Lima data for new employee data, which is what this URL is, and the target companies, and aggregate those two sets into one complete key personnel group. And then we would need — obviously there's existing employees in our database that doesn't need anything. Yeah, we might have a list of emails potentially and not the employees, and we can use another Lima data endpoint for that. Possibly.

The other piece is how do we get board members, how do we get advisors, and I think this is gonna have to be a scraping tool. Are you familiar with Exa or Firecrawl? Firecrawl, yeah. I assume one of those two tools might be what we're looking for here, at least for gathering. I think we would probably want to identify the sources that we want or need.

So like, I think the idea is — let's see, does Greptile have a — I assume Greptile's funded, right? I would assume so. Are they YC? Yeah, they're YC. So like, if they are — do they not have a team page? Yeah, I mean, that's definitely the traditional company will have that, but some of these SaaS startups, I think we'll have a mixed bag.

Okay so let's see — Postman.com — but that's easy, founders have names. I do have a Firecrawl account if you want to know. No, no, fine. But the point being, we can get it from here.

Going back to here — we're not running crawlers ourselves. Yeah, I'm totally in support of that. And then I guess, would you mind adding one little action item below that where like, it would probably be worth researching potential consistent data sources that might be free that we could use. I guess I can add — I mean, Crunchbase has this. Yeah, Crunchbase does, the Crunchbase Pro does. Yeah, that's 25 to 50 grand to start.

This one — step when ranking, suggests local paths first, path matter, or on demand from the user. I also think if the match count is low, that we need to think about — which is, David doesn't think this will work in most large cities. Fair enough. New York, LA, SF, maybe Chicago or Houston. Yeah, I think the top 10 probably.

The smaller cities, shared colleges — which I know college isn't here — it's these cohort things. Yeah, I was just thinking of ways that we can sort of score. I can't remember if we're getting updated people data for locations. I know we're getting it, it comes through on the Lima endpoints for people, it just doesn't necessarily — I can't guarantee receiving it, to be honest. We might be —

So that's a thing. Stephanie works at Cloudflare or wherever, but she used to work in finance at Goldman and she has a ton of relationships there, blah blah blah. Location relevance, but I guess it could be sort of viewed in aggregate with other data points. Yeah, that makes sense. That's gonna be a challenging but interesting one.

Yeah, there's some good questions here. The NAICS is probably the best way to do the industry taxonomy decision. And you have to pay a lot of money to get the NAICS, but generally there's ways to get it for free too.

Ex-employer connectors — we already do this. It's not LinkedIn. Well, it is. Yeah, so there's nothing crazy about this one.

Find prior co-worker connectors — oh, okay, sorry, I'm like wait, that doesn't make sense. Yes, this is like when your prospects and your connectors worked together. This is when they work in the same company, maybe not together.

Shared cohort connectors is what we were just talking about — school, exec program, fellowship, which we're not going to see, military, accelerator. If they're a Y Combinator company, things like that. And Lima may help for known programs like YC or Endeavor.

I gotta imagine that Y Combinator has a list of all of their companies on it? Oh, you probably want to — yeah, I would not go to that site. Just a question. Questionable things go in there. Yeah, okay, so there's a startup directory now. Questionable whether it's sourceable — like not scriptable — but I bet it's very scriptable. You think? Yeah, you would have to browse it non-headlessly and use the Claude for Chrome extension, but it's scriptable. Finite in length. Oh yeah, maybe 500 startups away. But there's a bunch of them. They rebranded as 500 Global.

In any way, the other piece is most investment firms have portfolio sites. So this is Lerer Hippeau, the one I'm familiar with — like it has information on it. I'm sure they got someone like that. Yeah, we definitely need to scrape those to start up with them too. The fact that they literally scraped their portfolio pages. Yeah, they're curious like a16z and stuff.

Given the list is sort of — what the fuck am I watching here? The price of RAM going up? Yeah. That's not what I'm watching. I like that they have a team page. Yeah, see they have one too. And the question is, okay, it is HTML so it should be easy enough to scrape. That's good. Nice. I don't know why they wouldn't block bots on these sites specifically. I think they want to index these pages, but yeah.

Shared investor paths — also cap table data is unsourced. Yes. Crunchbase for funding and then — I don't think Lima is going to give us anything. What about, not Crunchbase, well — the startup was Pitchbook, found out.

One thing that, because we did this whole exploration like two or three years ago, we were trying to think of what are non-LinkedIn options for data. And we've just discussed a bunch of them. But one that I always thought was interesting that we never really got to fully explore was tech partners — like, you know, who do you go to for your cloud infra, your tooling needs, what have you. And you can use the site builtwith.com for SaaS companies in particular to identify potential companies that have a real existing business relationship. That's a really good point.

Yeah, I don't know that that top one actually matched. I mean, it's interesting, because the problem in some ways with this is it's really good at catching all the analytics stuff. It's also really expensive to get the full premium data, but I do like the idea of being able to identify these potential tech partners as real established relationships. You can sort of deduce where a salesperson might know a leader in a certain department.

Yeah, I see. And this is through the website, but I bet if there's an API, it's worth it. And there is an API — if they're charging that much, right? There's a lot of crazy stuff here. Trends, profit, DPI, trust API. I mean, I guess I can see from some of the examples like you can look up Hotels Combined and see the relationships. Even builtwith.com — but it feels like the documentation reads as though it's like, I add this on my site and it tracks this stuff. It's worth exploring, it's definitely worth exploring, but it seems a little confusing.

Okay, I think it was the shared investor stuff. It's probably not investor path as much as it is shared cohort. I can't believe that isn't a connector role, come to think of it.

Okay, this seems to be filling in. So this is heading to the — my plan is to hand this back to Claude in a few minutes and have it update with our notes from our conversation. Nice. And then create tickets from it.

Explore third-degree paths. Yeah, this is continuous. I don't think there's any external sources for this. It's just data we have. Do we want to make it such that we can expose third degree in the sense of defining in the system what a third degree is?

Well, good news is we have an amazing front-end developer, just not a person here, who's already taking account of that. Third degrees are in the expanded network, the extended network. They're unverified. They go along with unverified signals. They're like, holy shit, we don't know anybody at this company, but we kind of know this guy who used to date this guy's sister — you know, those kind of relationships. So they'll show up down here in that they're not really part of the relationship map. As I was describing, these up here are almost guarantees. Down here, these aren't matches — these are chances. We're spitting at the roulette wheel.

Yeah, a little bit. Cool.

And then Path to Person — I don't think this should be external at all. I think it should be like — oh, I see. This was supposed to be you could add a person from the page. David had this idea at one point that you could be like, I want to add their CEO, and add the CEO to the relationship deal team. Because I want to add somebody to this that I want to know.

Because what I've been framing this as is — Rel.ai through here is just search. I want to find a person. Relationship mapping is I want to sell into this company. It's a broader type of search and it's an autonomous search. It's broader, it's autonomous, it's always kind of moving, and it's more nuanced. Whereas Rel.ai — the only time you would be manually adding people to it, potentially, I guess the tool is reasonable.

Yeah, and again, we don't have to do it all at once. We never have to build this one if we just never get there. I think that tool is probably all internal at this point, unless we begin to start using and aggregating and scoring with external data sources too. But that might be totally out of scope. The closest one we have to a plain natural language search for a person's name, no? That's probably true.

Yeah, but I think the problem there is that you're scoring — let me say it this way. Do you want to score on data we have in the database, or do you want to score on data that's in the database and also might be out there? It just creates a lot more complexity, especially if what we can do is pull their data into our database. And then it's just a closed set of data. The problem and challenge is keeping that data up to date, but we're already improving our ability to do that.

Yeah. Actually, I have it paused right now, but it wasn't paused this weekend. Because we're going through — now we're cranking. It's actually moving real fast. And actually, watch this. This is actually cool. I did this earlier and I was like, oh, that's so cool. If I pause the enrichment queue and once we only have — these will go in a minute. The Chrome extension threads will start tearing through them. Like 60 a minute. That actually doesn't seem as high. I think it was over 100. Yeah, but I mean with 30 threads going, I can imagine. Yeah, so right now we're probably about 25 threads I would bet. Yeah, look at that.

Hot dang, yeah, see that's what I was trying to get at — create lots and lots and lots of jobs, because at the end of the day we can spread out the burn, but we got to get the initial holdup out of the queue.

Yeah, because if an email comes in right now, it's going to wait until... well, actually, no, we moved those two notifications. But at the time, it would wait till the end, and that was a problem actually. Yeah, there were all sorts of problems with having blocking jobs in the queue with only four open threads.

Yeah, the thread thing. I mean, we just do "get Salesforce" now because it sounds like that's the shows of weeks above a couple of minutes. I'll be good.

Oh, these are... yeah, I'm surprised those are stuck. No, they're not stuck. They generally take seven or eight minutes, but I'm still surprised they're taking that long. I still think they're stuck. Analytics that could be improved — I did run some analysis and Claude's analysis was "make your views materialized views." Okay, if only we thought of that. Yeah, so there might be some there.

Okay, back to the task at hand. Wait, why are there... start. Yeah, see, default's already done. Yay! Thanks, okay, back to our work, Cams. Getting me all distracted.

Okay, so then I don't think there's anything too complex about these, and I don't know if there's much discussion because the targeting ones are all internal. Scoring, number of this decision point, ML later — I don't think that's right. I think it's all based on rules.

Barify relationship strength. Okay, so this is a... this is the thing, this tool just literally says "hey, what's your strength," right? Need to just review. And then... I love when Kwonk gets all hoity-toity. This is the IP. I don't think that's true. It's funny how they think we're machine learning engineers too. Yeah. What, you think this is a swarm?

And then engagement. Solution offers for help, build. Yeah, we're gonna — so this is where these things are also like, actually, this should be an engagement. Thank you. Even though it's in targeting, it should be here actually.

For what it's worth, and I've talked to you quite a bit about this, this stuff here in engagement is what I think we should be calling the concierge. Everything else should be like the relationship agent, but the concierge is just the email and communication back and forth. You know, honestly think of it like a concierge at a hotel. It's like a little groveling, you know, knows what you want before you know it, very supportive, as opposed to the other stuff. And just focus the marketing in that regard.

But all these are in there. Intelligence is just additional stuff. And honestly, this is all stuff like — I would say research started company is great. It's probably Lima, because Lima already has a lot of this data under the enriched company data.

So I'm kind of with the opinion that all three of those are probably the least — or would be the most difficult to build and deliver some high-impact, high-value tool versus the others, where the gains for the users would be tangible, measurable, and clear. Whereas this is sort of open-ended and it's very difficult to determine, at least at this stage, where the actionable insights would come from.

Yes, I agree. And so that's also why we don't need to do anything on these yet.

I'm going to pass this back to Claude. And then are you cool with me making a ticket for each one and putting it in a project? And then I'll start having Claude create tickets. I really, really, really like that.

I want to try and have — I really, you know, the connector dashboard, the request for stuff, and all the other stuff I worked on — I want to try and break these things down. Did I show it to you? The design system? Yeah, like the kind of design system. This thing? Yeah, you showed them that, or at least the intro. Yeah, I'm going to start trying to continue breaking it down so that we can try and just implement it consistently throughout the application. Absolutely.

And then the other piece, which is probably stupid and dumb, but I want to try to use Playwright for testing on this, because we're getting further and further away from the Rails stuff and we said we can't do it well today. That's kind of true but not as much anymore. So I think having the front end kind of broken out would be really useful. The fact that our front end's separate will make it easier. Yeah, definitely. And I mean, there's probably an argument for test coverage and stuff. I'm just thinking long term.

But yeah, so we used to use Puppeteer — not Playwright, Puppeteer — or, you know it, but like if they were constantly matching up or rendering different, it was a really cool idea in theory, but in practice it was not a lot of payoff. Not a lot of juice for the squeeze. But if we're taking the layouts apart and componentizing it a little bit, like a combination of unit tests — Jest or Vitest — and then browser automation tests in combination, if we want to do full front-end testing, then it does at least get easier to generate some tests. And those can be tests now.

I mean, I think that you have the source code and we've got our agentic coding tools, and we have it sort of like identified for browser automation using the Chrome DevTools CLI tool. And it could basically predefine every single browser automation test before even writing a single line of code and can validate it. And there's a Playwright MCP — I believe Chrome DevTools uses Puppeteer under the hood. So I think this is the year of browser automation technology.

Yeah, it's been a long time coming. Okay, so I'm gonna put this into tickets now. I'm gonna get going on that, and then tomorrow after stand-up, why don't we plan on talking?

OK, let's start kicking off a separate branch to keep it up to date. So I'm thinking a feature branch, and then you and I branch off that feature branch. Go take that time. Yeah, be real aggressive with constraining ourselves to follow the rules, so to speak. Like, you're gonna work on these first five things, I'm gonna work on these things. And when those things are done, we're going to pause, we're going to merge, we're going to relax. And I bet we can get something out by the end of next week that is at least office drop and relationship leads — which, to be clear, David has said if we have that, he's like, "I want to go and sell this immediately."

Okay, for router validation — my intent is to start writing this for the APIs. It's a bit of an API contract there. Oh, sweet. Yeah. So cool. Awesome. And that means we can also, by doing all this, get rid of all the introduction and all the rel leads — well, they'll still exist, but all the old APIs. And that gives us more that we can start moving other stuff over to. Over time it'll be great to reduce some of the clutter in that. Yeah, I agree. That's clever.

Ooh, what is it, 6.1? Yeah, which is like end of life a year ago. We're well past end of life, but even Rails 7 is about to be end of life this year. Yeah, Rails 7.1 and 7.0 are only on critical security updates. 7.2 is on stability, but 7.2 only runs — I think because of the Hotwire stuff? I think it's the Hotwire stuff. We also can only do this when we can remove all the front end, like the Webpacker stuff. Gosh, we're so close on that. We are. We'll get there in the next few months. We'll definitely get there.

And then we can switch to Go, right? Yes, and then let's create microservices. I mean, at that point my assumption is that CK will come back and take that over. So thanks for keeping my seat warm, Michael. I've thought — I'll be honest, and this is a little petty — but I've thought about that from time to time, like when Matt's logged in and seen the new front end. I'm like, I wonder if CK ever is like, you know, what are you guys doing over there? It's petty, but it is a thing. Sure, it's crazy. Crosses their mind.

Okay, I'm gonna take 10 minutes and then I'm gonna respond to David's post and stuff. Please don't, because I'll explain what's going on with Salesforce. I'll just say like, it's been held up, but it did run this weekend, so it's very likely that there's some targeting something or other going on, or a misunderstanding. We can dive in. Dave wants me to go over. Okay. Copy that.

Okie dokie. Thank you for your work today. I'll send this back over to you. Well, you'll start seeing tickets get created soon. Okay? Yep. Okay. Sounds like a plan.

**System Audio:**
Those bands zipping around, I'm doing shit. Why would you fucking say that, man? You don't say that because now you're going to make the universe give that to us. It's like saying the word yips.

Yeah, I saw some H-27s, but they're actually on the fucking V2 dashboard. So the quicker we can get rid of the old V2s, the better. Yep, we'll get rid of it soon.

And then I'm sorry about the invite bullshit again. I'm going to try to get us started on this agent stuff next day or two, and then I'm just going to tell David we need to take a week. I put together a couple tickets for completely removing all the IAM stuff from Super Connector and just rolling out complete Work OS integration invitations and all that. If we could just do that, that stops one of the largest internal confusion points for him.

Yeah. Do you want to finish your email? I'm at the point where I'm now pasting all my emails into Claude being like, am I coming across weird here? Yeah. Okay. I thought I put this email in. I just realized the stupid point.

The other thing is, you know the whole Claude em-dash thing? Yeah. I have been using em-dashes since high school. Like, I really care about em-dashes, really care about the proper use of em-dashes. And so it's hard because I now have to take them out. It's interesting because — I don't mind the un-dash. I mind it because I want to change my emails. I don't like having to change my style of typing or writing, but I don't mind it. But I'd be really pissed if it started replacing em-dashes with comma space comma, which obviously it would never do that, but that's what you just said at the same time. I'd be super pissed. I'm like, you're clearly wasting tokens, like you don't need to do that.

Has Claude started swearing anymore? Because mine keeps going — shit, I messed that up. Interesting. Maybe, yeah, that can make sense.

I know you have hard feelings — or have strong feelings about it. I'm still incredibly happy with Claude. Like, Todd's comment this morning at stand-up, I was like, it's like a small thing to get upset about, I guess. Like, sure, but I'm not buying it. Yeah, but still, okay. You probably can't tell him that, actually, but it's cute that you think so.

Okay, so I have to be careful that I don't make jokes like this too much for people who are going to think I'm not good at my job, but — doing what I normally do, not my job, and just having Claude do it. I took our conversation and post and had Claude read it, plus my handover docs I generated from the prototype development. And so what I really had to do, and I'll share this with you, though I don't know that we want to keep this in the document — I think we want to solidify this and then make each one of these into a Linear ticket in a project. But what I had to do was just break down all of the things we had talked about, essentially, and go through them.

And so I don't know, I'm starting to really like this idea of — given how fast the connector dashboard works when Sidekick is actually running — the Elasticsearch, I think. I'm sorry, I really like it's the way.

Talking through them? Okay, let me — okay, I'm going to message you the link so you can pop it up on your screen if you want.

Yeah, so, okay. Map account network is kind of continuous, you know. So it takes the idea that charter company ID is an odd way to describe it. I have a funny feeling. Yeah, but sometimes Granola does. I must mumble or something. Thank you. Yeah, I've thought a lot about getting a new mic instead of this thing, but we'll see.

Okay. So it takes the account ID, target ID, and then it provides the output. Oh, I'm on the wrong one. That's right. It takes the target ID, account ID, and returns relationship IDs. So this is just continuous rerunning on a cadence, you know, on relationships.

What the hell, did you send me a web page? Okay. Yeah, and so this is coming from the prospect search index probably, or something of that nature. Makes sense, right?

Now, Kiefer Snell — this has got it wrong, I think. I think that this has to come from Lima data. So something like prospect company or prospect employees. Cool. I agree with that.

And I think advisors? So we would need new Lima data for new employee data, which is what this URL is. And then we would need — obviously there's existing, that doesn't need anything, possibly. The other piece is how do we get board members? How do we get advisors? And I think that's going to have to be a scraping tool.

Give me a — so like, I think the idea is, let's see, does Greptile have a — I assume Greptile has money, right? Because if they — oh yeah, are they YC? Yeah, that might be it. If they are — team page. Yeah, okay. Let's see, postman.com. Okay, that's not useful at all. But yes. No, no, fine. But the point being, we could get it from here.

So going back to here — I'm just going to be blunt, we're not running crawlers ourselves. Okay.

And then, yeah, I mean Crunchbase has this. 50 grand to start.

So then, this one — agent reasoning step when ranking suggests local paths, first path of matter, or on-demand from the user. I also think there's some notes here that we need to think about, which is David doesn't think this will work in like most large cities. New York, LA, SF, maybe Chicago, probably Houston. But will work.

College is in here. I cannot remember if we're getting updated people data for locations. I know we're getting it. It comes through on the Lima endpoints for people. It just doesn't necessarily — yeah, so that's the thing.

You're aware, this is like the Stephanie Cohen idea, which is — but she used to work in finance at Goldman. She has a ton of relationships there, blah blah blah. Yeah. Questions here.

The NAICS is probably the best way to do the industry taxonomy decision. You have to pay a lot of money to get the NAICS, but generally there's ways to get it for free too.

Ex-employer connectors — we already do this. It's not LinkedIn. Well, it is. Yeah, so there's nothing crazy about this one.

Find prior co-worker connectors — okay, sorry, I'm like wait, that doesn't make sense. Yes, this is like when your prospects and your connectors worked together. This is when they worked at the same company, maybe not together.

Shared cohort connectors is what we were just talking about. We get education from LinkedIn and LIMA. May cool exec program, fellowship — which we're not going to see — military, accelerator if they're a Y Combinator company, things like that.

Help for known programs like YC or Endeavor — I gotta imagine that Y Combinator has a list of all of their companies on it. Oh, you probably want to, you know, I should start sourcing correctly. There's some questionable things going on there. Yeah, okay, so there's a start directory now — questionable whether it's sourceable, like scriptable, but yeah. The other, like, isn't 500 Startups one of them? Oh yeah, maybe. 500 Startups — there's a bunch of them. Oh, they rebranded as a VC fund called 500 Global.

In any way, the other piece is most investment firms have portfolio sites. So we could also generate a list of VCs and scrape their portfolio pages. Great douchebag startup funds too. The fact that they literally — oh boy, what the fuck am I watching here? Yeah, that's not what I wanted. Though I like that they have a team page. Yeah, see, they have one too. And the question is, okay, it is HTML so it should be easy enough to scrape.

Cool. Shared investor pass also fits right into this — is unsourced. Yes. I don't think LIMA is going to get us anything. Yeah. I thought, not Crunchbase — well, the startup site was called WellFound now, the AngelList went away, right? Oh yeah. It's just a fund administration platform though.

That's a really good point. Yeah, I don't know that that top one actually matched. I mean, it's interesting because the problem in some ways with this is like it's really good at catching all the analytics stuff, but like... Yeah, I see. And it's probably — this is through the website, but I bet you if there's an API requirement, I bet you — no, I mean, frankly, yeah. There's a lot of crazy stuff here, but like... What's the... Trends, Product API, Trust API. Okay, great. But how much does that cost? Yeah. It's worth exploring. It's definitely worth exploring, but it seems a little confusing. I don't think that goes under shared investors though.

Okay, this needs to be filled in. Okay, so this is heading two, this should be heading two. I want to hand this back to Google or to Claude in a few minutes and have it update with our notes from our conversation and then create tickets from it.

Explore third-degree paths — yeah, this is continuous. I don't think there's any external sources for this. It's just data we have. Because if you — well, good news is we have an amazing front-end developer/designer person here who's already taken account of that. Third degrees are in the expanded network, they're unverified. They go along with unverified signals. They're like, holy shit, we don't know anybody in this company, but we kind of know this guy who used to date this guy's sister, you know, kind of relationships. So they'll show up down here in that. They're not really part of the relationship map there. As I was describing, like, these are up here at offers for health. Yeah, a little bit, yeah. Cool.

And then path to named person is — I don't think this should be external at all. I think it should be like — oh, I see. This was supposed to be, you could add a person first. David had this idea at one point that you could be like, I want to add their CEO. Yeah, or like right here, put him in the relationship deal team. I want to add somebody to this that I want to know.

Because what I've been framing this as is Rel.ai through here is just search — I want to find a person. Relationship mapping is I want to sell into this company. You know, it's a broader type of search and it's an autonomous search. It's broader, it's autonomous, it's more nuanced, whereas relationship AI is find people who look like this, and I can decide how to interact with them. So adding something to the relationship map is like the only time that you would manually add people to it potentially.

Yeah, and again, we don't have to do it all at once. We never have to build this one if we just never get there. That's probably true.

Yeah, but I think the problem there is that you're scoring — let me say it this way. Do you want to score on data we have in the database? Or do you want to score on data that's in the database and also might be out there? It just creates a lot more complexity, especially if what we can do is pull their data into our database. And then it's just like, this is a closed set of data. And then, you know, the problem and challenge there is keeping that data up to date, but we're already improving our ability to do that. Or aiming to improve our ability to do that.

Lots of — okay, I can't right now. David, asking about SFDC. I will follow up after. I mean, actually, I have it paused right now. It wasn't paused this weekend. Yeah, because we're going through. Yeah, no, it's actually moving real fast, which is a good sign. Pause the enrichment queue. But once all — like 60 a minute, but that actually doesn't seem as high. I think it was over 100. Yeah, so right now we're probably about 25 threads, I would bet. Yeah, look at that.

See, that's what I was trying to get at — create lots and lots and lots of jobs because at the end of the day, we can spread out the burn, but we got to get the initial hold up out of the queue. Because like if an email comes in right now it's going to wait until — well, actually no, we moved those two notifications. But at the time it would wait till the end, and that was a problem. Actually, that's yeah, the thread thing.

Yeah, I mean, main problem here — oh, actually, now it's going to pause us because these are going to take some time. They're not stuck. They generally take seven or eight minutes, but I'm still surprised they're taking that long. I still think there's some analytics that can be improved. I did run an analysis and Claude's analysis was "make your views materialized views." It's like, oh, okay. So there might need to be some improvement there, but okay.

Back to the task at hand. Wait, why are there — oh, there's only 30 because there's only supposed to be 30. Sorry. Thanks, okay. To our work camps. Get me all distracted.

Okay, so then I don't think there's anything too complex about these, and I don't know that there's much discussion because they're just — the targeting ones are all internal. Scoring is a decision point, ML later. I don't think that's right. I think it's all based on Bearfighter relationship strength. Okay. This is the thing. This tool just literally says, hey, what's your strength, right? And then I love when Claude gets all hoity-toity. This is the IP. I don't think that's true, actually, but okay.

Yeah, solicit offers for help. Yeah, we're going to — this is where these things are also where, actually, and this should be an engagement too, even though it's in targeting. Should be here, actually.

For what it's worth, and I've talked quite a bit about this — this stuff here in engagement is what I think we should be calling the concierge. Everything else should be like the relationship agent, but the concierge is just the email and communication back and forth. You know, just honestly think of it like a concierge at a hotel. A little groveling, a little knows what you want before you know it, you know, very supportive as opposed to the other stuff. And just focus the marketing in that regard.

But all these are in there. Intelligence is just additional stuff. And honestly, this is all stuff like — I would say Research Start Company is great. It's probably Lima because Lima already has a lot of this data under the Enriched Company data. So just start thinking about them, then for each one putting in a project and then I'll start having to create tickets.

I really, really, really want to try and have — I really, you know, the connector dashboard, the requester stuff and all the other stuff I worked on, like, I want to try and break these things down. Did I show it to you, this piece? Yeah, I like the kind of design system. This thing. So I'm going to start trying to continue breaking down so that we can try and just...

And then the other piece, which is probably stupid and dumb, but I want to try to use Playwright for testing on this. Because we're getting further and further away from the Rails stuff. And we said we can't do it while it's in Rails. And that's kind of true, but not as much anymore. So I think having the front end kind of broken out would be really useful. Make it easier.

Yeah, one of them was — I can't remember the name of it, but like you would render the same component a thousand times with a lot of different data and you would look to see if they were constantly matching up or what. Rendering different. It was a really cool idea in theory, but in practice it was like not a lot of payoff, you know, not a lot of juice for the squeeze. But if we're taking the layouts apart and componentizing it a little bit like this, then it does at least get easier to generate some tests. And those can be unit tests. Yeah. Okay.

This code going — what is our process? How do we keep this separate from main but also keep it up to date with main? So I'm thinking a feature branch, and then you and I branch off that feature branch. Yeah, be real aggressive with constraining ourselves to like follow the rules, so to speak. And like you work on — you're gonna work on these first five things. We're going to start working from. And he said, if we have that, he's like, I want to go out and fucking tell us immediately. Okay.

Yeah, I also really liked what we were doing, what I asked you to do on the form — like the forms for parameter validation — is to start writing this for the APIs as a bit of an API contract there. So, cool. And that means we can also, by doing all this, this will get rid of like all the introduction, all the rel leads — well, they'll still exist, but all the old APIs. And that gives more that we can start moving other stuff over to over time.

Agree. 7 in the summer. Yeah, because we're on, you know, 6.2 — what is it, 6.2.1? 6.2.4? We're well past end of life, but even Rails 7 is about to be end of life this year. Yeah, Rails 7.1 and 7.0 are only on critical security update. 7.2 is still on critical security update. Because it's really hard to jump from six to eight. Yeah, you really want to go six to seven then to eight. All the front end, like the webpacker stuff — we're still... we are. We'll get there in the next few months. We'll definitely get there. I feel confident.

And then we can switch to Go, right? That was the plan all along. I mean, at that point, my assumption is that CK will come back and take that Go work because — I've thought, I'll be honest, and this is a little petty, but I've thought about that from time to time.

I wonder if Matt's logged in and seen the new front ends, or if CK ever is like, "What are you guys doing over there?" It's petty, but it is a thing.

Okay, I'm going to take 10 minutes and then I'm going to respond to David's poster and stuff. Please don't, because I'll explain what's going on with Salesforce. I'll just say it's been held up, but it did run this weekend, so it's very like...

And then move around four o'clock because David wants to meet to go over something at 3:45 or four o'clock.

Okay. Okey dokey. Thank you for your work today. I'll send this back over to you. Well, you'll start seeing tickets get integrated soon, okay?

Cool. Thank you. Talk soon. Bye.

Cameron and Michael held a working session to review and refine a breakdown of the relationship mapping / agent feature set for SmallWorld's platform. They walked through each proposed tool — map account network, key personnel identification, shared cohort connectors, third-degree paths, and engagement/concierge features — identifying data sources (Lima, Crunchbase, BuiltWith, VC portfolio scraping) and scoping complexity. Cameron plans to feed their conversation notes back to Claude to generate Linear tickets in a project. They also discussed near-term priorities: removing IAM from SuperConnector in favor of full WorkOS integration, a feature-branch strategy for parallel development, introducing Playwright for front-end testing, API contract validation via router schemas, and the eventual Rails 6→7 upgrade path.