They agreed the next Kevin step is a direct budget conversation before any numbers or proposal are put in front of him, with Andrew taking that call as soon as he is comfortable. Andrew also agreed to review Danny’s Kevin materials and give feedback, and said he wants to look further into the recall/memory tooling Danny described. The discussion on presentation left a preference for leading with the top wins and keeping fuller detail below the fold, but no final format change was settled. Danny argued that meeting summaries from Granola and similar notetakers are not reliable enough as the primary record, especially without diarization and when deeper recall is needed. He described using a mix bread instance with raw transcript feeds, embeddings, and MCP/CLI access to support broader recall across agents; Andrew found this relevant and worth investigating, though he has not yet felt a strong personal pain around it. On Kevin, Danny saw a need for a larger scoped proof and likely retainer/upskilling rather than a low-end flex-capacity arrangement, while the budget, margins, and commercial structure remained uncertain. Andrew floated a straight retainer, a reduced first month, or outcome-tied compensation, but treated outcome-based comp as probably too cute for now; Danny favored proving value through one focused standalone slice. Andrew found Danny’s artifacts generally good but possibly verbose and said he would dig in and provide feedback.
Right? Like,
I've looked at it not reliably, not in ways that are like interesting like like they really that's the piece they want to own besides like they want to own all of your beating intelligence and make it free for most people, you know the five year play I they they restrict some of the like it's at a point where the amount of like tooling you build around granola means you should just build a super lightweight because the, the open source, the state of open source with respect to like the tooling that was doing the passive input output capture is
now just totally solved. And then outside of that the diarization there's a handful of open source tools and paid tools if it's more interesting that can sit between that local layer and do the diarizing. And like
When you say diarizing, what do you mean?
I mean like there's three of us in the call. Danny said this then, Andrew said this then and Bob said this then that is like people are, here's what people are doing. People are relying on, on the provided MCP calls that granola furnishes and what is that default for granola get summary. And so people are lit and every note taker does this. People are living off the summary. The summary almost always is not the full picture. Right. Because now it's going to depend on which model or models were involved in this summarization process. Okay, like is there web context? Was there enrichment? Yes or no? Great, back to the transcript. It may not matter if there's no diarization. If it's granola, it's me them in a one on one perfect, fantastic.
Outside of that probably noisy like and I'm talking you will still get 85% 90% but I'm talking about the I'm deep in shit and I want to make sure this agent grabs that granola. I don't want to think about it like and so.
I mean, that makes a sense. That makes sense from your perspective there. I think, like I said, I'm really, I don't have, like, a hard and fast rule on this. Like, I'm not necessarily, like, craving every second of every thing. Context right now. Now that that may change. So it's, for me, it's really like, like, if I'm going to go and build my own tooling, especially if that tooling exists for sale, like, I want to have a reason to do it, which is what you've described. Right. Like, and so right now, I don't have that pain. Therefore, I haven't thought very strongly about it.
Sure. And I, I would posit you can't feel that pain until you're already running with the absence of this tooling in place.
Because to then change how you're doing, meeting intelligence capture or whatever, or to build on top of it like feels much worse in my experience as I've like had tons of great conversations but not gotten the context out of them when I was relying on a recording, a transcription and and so I just don't rely on them. I assume there is absolutely no recording. If there's something critical I'm working on, I'm taking notes as well. I hope for that to fall away. But how I'm also then managing this is like I've stood up a mix bread instance which allows me to aggregate and embed all of the context that I want in different data stores. In this sense I have a granola feed that's all of the granola calls, raw transcripts feed, feed into it. I have some notion embedding as well. And then the mixed spread MCP and CLI are made available for rapid recall recall across any agent. So like I don't even trust granola to be the place where I get what I know what happened last week in this call because candidly they are underpricing the out of the tool. I mean you have fable
Right?
in that drop down, you know, like they are just burning cash and it's going to hit harder I think than some people will be hit by ChatGPT becoming, you know, a lot more expensive. Like when this like you're meeting intelligence recall, like
that feels almost more consequential. So anyway, I've been trying, I've been slowly just kind of experimenting with this to like not only then get like rich context into agent runs and kind of have some of that analysis layer. It's just, it just feels stronger, safer to have like an embedded sense of meeting recall. Because my corpus is so large now that if I am doing a long running ask where I'm just how has my work evolved around this concept over the last three months or whatever? Like the time and inefficiency of a bunch of MCP calls like is almost always outweighed now by one call for a larger body of embedded context to then let the agent sift through what it needs. It's just like get to the thing, at least in my experience so far.
No, I mean, this. This. This makes sense to me as a. As. As tooling for, you know, your agent and, like, the body of info you wanted to work on, I think. Yeah. I mean, this is interesting. This is kind of, like, very relevant to some of the stuff I've been thinking about.
I was wondering. Yeah. So I wanted to bring that up today. It's something I'd implemented at Proaction towards the end and really liked the team really liked the idea. Their whole idea is around building primitives to let people build their own kind of recall and memory systems, not like buy my recall and memory system. So I tend to like and trust those kinds of guys a little bit more because they also had experience at open source. So. So anyway, something to keep on your radar.
Yeah. Cool, man. I'm gonna. This is. This is definitely something I want to look into more. Yeah. So what's. What's. Where. Where's. Where's your head at? Like, what's. Let's start there.
Kevin needs a lot of help. Because Kevin's goals aren't going to change. Some of them aren't his goals. So the company's goals and like those aren't. Those are unlikely to go in a let's do less direction. I would say just based on like they're getting those, getting. They're getting those little like immediate like light bulbs turning on, the fireworks are popping. Like there's, there's, there's emotion enough that they're going to keep pushing the question I'm still have no sense of how many dollars do they push with when they push, you know, like how. What are the margins on this business? It could be as. It could be really solid, but it could also suck.
it's. This may be wrong, but for whatever reason, when you said that my brain says they're in like 30% somewhere margins, that may be wrong, but, like, for whatever reason, that's what comes to my brain,
And top line revenue for that location, how many locations that staff? I mean, you know, it's.
pretty sure. Well, he's. He's not been specific, but he sort of alluded to at least a few million a year in top line revenue.
So a guy like that can be super well positioned to be like, yeah, 10 of our top line revenue for a year to make 25, the next five or whatever, you know, like duh. And if it's a duh situation then jumping to the retainer and be like, look, there are six projects. If we start do two parts of these in parallel. Like that means you'll have these systems you can start running and then we can build out these pieces. Like they want the repeatable model, but they can't get that unless they make an investment. And I just mean even like you've got a 20 Claude plan like so you haven't even started is what I was like
in the sense of like you have not to just put all of the data that your business might be sitting on through a cloud in the most inefficient way. You need a lot more than a $20.
Yeah, yeah, yeah. And so my thinking is, I think there it's, you know, it's like, what the broad steps. Like, you don't know you have a problem. You know you have a problem. You know, you have a problem and you want to do something about it. You know, you have a problem, you want to do something about it, and you've started doing something. Like, of those, I think they're at the. They know they have a problem and they want to do something about it. I don't know. And maybe this is. Maybe this is next steps on me is to, you know, just have a conversation with Kevin and say, hey, you know, we think there's a lot here. You know, here's some. Just broad thoughts on what we could even cover. But, yeah, all this, like, takes time and takes cost. Like, what do you think? Like, what. Like, are we talking, you know, 10 bucks a month here? Are we talking 10,000 bucks a month? Yeah. Are we talking, you know, 20,000 bucks? I'm like, what. What does this realistically look at? Like, and that probably. That. That's probably the next conversation that needs to be had because we just need to know, like, is there a budget here? And that's. I can. I can take that as a call to have if you want.
Yeah. I mean, I'm happy to be on that call. Like, I mean, it's just like. Yeah. Like, I think it's got to be, like, some combination of, like, reflection of deep understanding from provided context. And I've already got some artifacts around that I don't even want to show, like, a direction where it could go. If it's like, again, if fundamentally he's thinking, like, man, I give these guys two grand a month, and, like, I got flex capacity or whatever, you know? But if it's 20 grand a month, then it's like, yeah, you've got flex capacity. Like, yeah, like, let's go.
And that's what I'm saying. Like, I think I would say like the next call is just a person to person conversation. It's like, look, we think,
It.