The Foundation of an AI-Native Firm

There's an old idea that we inherit the wisdom of the generations before us, and that it's what lets us move through life a little wiser.

Twenty years of True Ventures is a lot of inherited wisdom to step into.  It hit me over time, as I grasped just how many pitches our team sees and how many recap write-ups we do in a single week - and then multiply that over twenty years. What we learn from all of those pitches and every single founder we meet is an invaluable layer of our team's documented collective intelligence, because True has a thriving writing culture.

This culture of deep documentation handed us the most interesting problem to solve: How do we turn our collective intelligence (both the documented and the hidden knowledge across our entire firm, our team, and an insanely talented founder community) into intelligence that's accessible team-wide? And how do we do it in a way that doesn't just feed our past back to us, but also aims us forward in a future-facing way so our decision-making can evolve?

These were some of the questions we set out to explore as we built True's Datalake, our central source of truth: a living archive of everything the firm knows, documented and hidden, that any model can query in real time to sharpen how we make decisions.

Here, I'll unpack how we built it over my first eight weeks at True Ventures — and how we rolled it out firm-wide to do everything from deep research, to finance operations tooling, to mining our networks with live agents so we can learn and meet with more AI builders and founders.

Where we started: context is king

In my first month at True, I was struck by the volume of stories held within the firm and in the memory of its team members. How much of what they know about the companies that have done well, and the founder archetypes behind them, is actually documented somewhere?

Surprisingly, a lot. That comes back to our writing culture rooted, and not just in individual inboxes. Our team has always “worked out loud,” at least internally, before it was a thing, sharing candid takes on pitches and the founders behind them in team-wide workspaces. That turned out to be wildly helpful when we architected True's Datalake.

So our first step was pulling together every knowledge base across the firm: P2 posts from the late 2000s and early 2010s where we documented our decision making processes, more recent post-pitch Basecamp and Notion write-ups, Granola meeting transcripts and summaries, email, Slack messages, community conversation summaries with insights gleaned from the room, and founder discussion threads.

We pulled all of this data into GCP buckets. We already ran on Google Workspace, so our email and calendar connectors were cleaner there, and BigQuery let us put twenty years of very different data behind one query layer. After that we normalized the data.

Every source had its own formatting, e.g., a Granola transcript, a 2009 P2 post, and a founder forum thread have almost nothing in common structurally. To account for that, we transformed everything into a curated set of tables with shared schemas, and resolved entities across all of it — using the help of AI and my teammate Mike’s knowledge of building distributed systems at scale from his time at Twitter.

That last part mattered more than we expected: the same founder might show up under three email addresses, a LinkedIn URL, and a nickname in Slack, and the lake is only useful if it knows those are one person. We keep the curated tables in BigQuery and mirror them into DuckDB, an in-memory database that makes queries feel instant.

Once this was all set in stone, we built a few MCPs we could call directly from any model of our choice. It was crazy: information that used to take hours, if not days, to pin down accurately was suddenly there in an instant. That changed the game for retrieval, and even for spotting trends across our history.

It still amazes me that all of a sudden we could visualize these stories: our trajectories, our growth, our mistakes. This datalake became the foundation for what came next.

→  Context is everything right now. If information is unique to your firm or startup, it's gold. Treat it as such.

Go ham: rapid build + integration mode

This phase of building the True Datalake kind of felt like "founder mode" for us. (Arguably, we are always in it at some level at True. It's where we all love to be.) And yes, it's fun to build tools, but they only matter if they're used, which means designing them to fit into existing workflows, or making the output so good that adoption is the obvious choice.

We could build 100 agents running on a cron and have only one be truly useful, especially for a smaller team. That forced us to answer real questions: how did we want to design an AI-centered flow, and what did it mean for us to be an AI-native firm?

Once we had our True Datalake as our foundational model, we could build all kinds of fun additional tools on top of it, and we went in every direction. Some of what we built:

• Agents that live right alongside us within our internal communication platforms
• A live copilot that sits in on diligence calls and can recall what a founder said two minutes ago
• A network atlas of over a thousand people and six thousand relationships — everyone True has ever worked with — built purely from this data
• A pre-meeting brief that lives inside Google Calendar as a Workspace app, so it fits the flows we already work in
• A self-healing loop: flag a wrong answer from the lake, and an agent files the ticket, writes the fix, and opens a pull request (the system literally repairs itself)
• Tools for our finance and platform teams, and now tools built to deliver as much value as we can to our founders.

And on top of it all, we let dozens of agents run in parallel to observe our data.

Revving the engine + continued refinement

Building fast and playing hard is exhilarating, but it leaves you with a pile of parts. The quieter, harder work comes next: figuring out how they fit together.

Individual parts are only as good as the way we put them together. Using AI well, we realized, means 100x-ing what we're already good at and where we do our best work, not necessarily inventing new workflows to get there. Even so, our 1.0 of True's Datalake still feels like a trial-and-error phase. We’re working as a team to bring out the best of what we know with AI.

Right now, our team is actively pulling insights from the datalake on a daily basis:

→ Mike is on it before, during, and after pitches — walking into a first call already knowing every touchpoint we've had with a founder, and asking sharper questions in the moment because the context is live.

→ Soul has been using it for deep research on companies we've seen before: how a market looked when we last evaluated it, what we believed then, and how that holds up against what we're seeing now.

→ Rohit's been using it to understand scale, for example sizing markets but also the tech in real time to understand not just the business but our own history in the space.

Everyone at the firm is using it differently, and it's been growing organically because of that. And this is just the beginning.

This is just the start

True's Datalake is our 1.0, and there's so much more we'd like to build around it: more AI workflows and pillars that can turn all this intelligence into boosters for how we actually work every day. I'll keep sharing what we're building at True as it comes to life, so follow along or reach out if you want to see where this goes.

Keerthi Nalabotu is True’s AI Builder in Residence, role focused on creating AI workflows, automations, and tools that help our investment team accelerate and sharpen. Follow her on X or LinkedIn.