Why We Built a VC Firm Like an Intelligence Agency: SIGINT

SIGINT (Signals Intelligence) is probably my favorite intelligence practice to say aloud - SIGINT, sounds like signet (engraved ring) or cygnet (baby swan). But even more than the phonetic play, SIGINT reflects the very real machinery that we’re building at Theory that powers all of our intelligence practices. While we generate incredible learnings through all of our intelligence work, SIGINT carries the critical burden of making those insights actionable, building both the infrastructure as well as applications to ensure the entire team is delighted to engage with our systems. These systems turn the everyday work of all of our teams into a compounding machine that makes tomorrow better.

SIGINT encompasses all of our internal and external data sources that inform everything we do as a firm, from initial sourcing, our diligence, our portfolio support, and ultimately to our brand work. The research that drives our investing needs to encompass all potential signals we gather, whether that's from in-person events or blog posts or podcasts we listen to, as well as the more structured data and the expert calls. Our SIGINT team builds the software to accomplish this: architectural decisions, agent orchestration, inference optimization, tool specification, and so on.

It's not enough to make our data accessible, but we need to make it frictionless to collect, grow, and leverage, as opposed to requiring heavy manual steps, which leads to lossiness. That creates a set of hard problems: resolving the same company or person across dozens of messy sources, changing domains, and more, turning unstructured conversations into something you can query, and building evals rigorous enough that we trust the outputs we make decisions on.

Leading this work is Bryan Bischof, our Head of AI, whom you might know for his colorful attire and sometimes even more colorful X posts. Bryan has led data science and applied AI teams at multiple unicorns, such as Hex, Weights and Biases, Stitch Fix to name a few, and he’s now building our software at the frontier of the AI application layer, adopting and experimenting with new technologies that transcend the engineering work itself to inform our investment theses. For example, our SIGINT team adopted and built our company-brain-as-internal-MCP nearly nine months before it was the obvious route for enterprises.

SIGINT Team

• Adam Conway, Senior AI Engineer: Tech Lead at Modern Health, Founding Engineer at Kismet, Architect at Stitch Fix, Mobile Lead at Blue Bottle Coffee.
• Melanie Rosenberg, Senior AI Engineer: founding AI engineer at Town, where she built AI agents for tax automation. Before that, she helped launch Speechlab, a voice-cloning startup out of Andrew Ng’s AI Fund, and led ML at Tellus You Care, building radar-based health monitoring.
• Panth Bhavsar, AI Engineer: Waterloo computer engineer who supported VC fund operations at Carta and built his own skincare app from zero to one, so he’s seen both sides of the founder–investor table.
• Mischa Lamoureux, AI Engineer: Waterloo engineer who worked on Tesla’s Cybertruck software, coauthor of last year’s Hackathon, semantic.art, and is our external facing engineer.

The SIGINT team is not only building internally for Theory. They’re at the table for our deal diligences, evaluating technical roadmaps and weighing claims. After we invest, they’re who our founders call about evals, agent infrastructure, or hiring the best engineers. When a Theory engineer reviews your stack, they’ve shipped in production themselves.

While the team has been forthcoming in capturing how we think about some of the biggest challenges in AI and the technical solutions we're leveraging (e.g. how to manage rolling back agent actions, how to give agents guidance in architecture tradeoffs, how to develop and manage skills as an enterprise), I thought it might be most illustrative to go through a real-life case of our systems in practice.

In the process of building a new feature, Adam receives an ask from Spencer. Adam uses an LLM to generate a response and then sends that back to Spencer, who then shares that response with his LLM for further review. This back and forth highlighted the emergence of a technical need: teammates aren’t collaborating one-to-one, but instead agents are in the conversation, too. Existing chat tools aren’t built for this.

Adam identified this gap and built a thesis around it. He leveraged our systems to pull every relevant conversation that touched this space, identify buyers, build a competitive analysis, & size the market, all in a matter of hours, compared to work that would take an analyst weeks. Because no clear player existed yet, he went one step further to identify who in our network might be best positioned to found this company. He was able to complete all of this investment work while at full capacity on engineering projects because of the deeply connected systems that we nurture and develop.

This demonstrates the leverage of the custom software SIGINT is building, that a software engineer can rapidly ramp up investor-grade research in a thesis he's lived firsthand.

We’re often asked, how much is left to build? Truly, we’re still in the earliest innings. With new technical challenges constantly afoot, we’re building faster than ever, and every new primitive we develop unlocks the next wave of companies.