Building Colabra

Our team has built and deployed software for PE customers that analyzes tens of thousands of financial and legal documents across a data room, finds inconsistencies and flags deal risks with supporting sources.

In a published customer testimonial, Ray Nexus reports reducing its diligence timeline from three to four months to about one month, on deals ranging from $5 million to $105 million.

An earlier version asked users to assign individual tasks to an AI collaborator. We changed it to review the data room as a whole, with findings and follow-up work attached to the source documents. Colabra could take on more of the review without requiring the user to assign each task.

The product overview and documentation describe the system. Through its MCP connection, customers can use an external AI client to work with project evidence they have permission to access.

We started Colabra as an AI company focused on scientific R&D. My cofounder and I had been experimenting with GPT-3 together since 2020. Researchers were pulling information from paper lab notebooks, machine outputs, and external datasets, then sharing it through SharePoint and monthly meetings. We built document handling, search, permissions, and collaboration around that work.

I went to Kyoto on a day's notice to meet Nobel laureate Shinya Yamanaka's team and secured a paid contract. I followed through with repeated visits to the lab in San Francisco. The team used Colabra every day to run that lab from Japan, reviewing experiments and directing work with the data, permissions, and discussion in one place.

Nine days after GPT-4 launched in March 2023, we had it reasoning over real experimental data from Colabra while preserving each user's permissions. The prototype ran in a Jupyter notebook, using the document handling and access controls we had built for research teams.

Later that year, we added hybrid vector search to support questions such as "Find experiments that struggled with E. coli growth."

Selling collaboration software proved difficult: smaller labs had limited budgets, and larger organizations didn't always have a clear buyer for a tool used across departments.

Conversations with investors buying scientific-software companies introduced me to M&A data rooms. The buyer needed to analyze the documents, reconcile the evidence, and track unanswered questions. Much of the infrastructure we had built for research could support that work. We explored several adjacent markets and chose buy-side diligence because the review had a clearer owner, a deadline, and an existing budget.

I flew to meet prospective customers, took red-eyes to conferences, and built relationships across industries that led to warm introductions. I used those conversations to refine the product and our pitch.

Alongside direct sales to labs, I contacted about twenty podcasts and twenty newsletters to reach researchers through the publications they followed. We couldn't offer an acquisition payment, but we could take over editing, publishing, distribution, and guest booking. That led to our partnership with Talking Biotech.

Before an episode, I spoke with a prospective guest about their research, how their team collaborated, and what they were working on. Those pre-interviews helped me prepare the show and understand how potential customers worked.

When we moved into diligence, deal frequency became an early qualification question. Some smaller PE firms were interested but completed too few transactions to use Colabra regularly.

For a firm with enough deal activity, I work out who will use the findings and what they need to evaluate. I've proposed tests on historical deals so buyers could compare our findings with their existing review. We'd agree on the questions before the test and examine where the software missed something or reached a different conclusion.

I build tools that draw on earlier conversations to help me prepare for meetings and review follow-ups. I describe that work, along with the evaluation proposals, in Technical GTM for AI products.

We use synthetic data rooms with known issues to test which parts of a diligence review benefit from Colabra as models improve. We check what the AI finds, whether the sources support its findings, and where it makes mistakes. We can then decide what to improve in Colabra and what to leave to the model without disclosing a customer's documents.

In our July 2026 internal benchmark, AI using Colabra scored up to 31% higher on output quality than the same AI using a leading data room's public document-retrieval tools.

We gave GPT-5.6 the same eleven diligence workflows, 1,799 synthetic files, questions, prompts, and time budget in both configurations. Only access to the files changed. Colabra led all eleven workflows against the data room's public MCP connection, with an average output-quality score of 8.30 versus 7.30 out of ten. The 31% figure is the largest gain on an individual workflow.

These scores describe the tested setup on synthetic documents and a public MCP connection, rather than the provider's internal AI systems. We also run separate comparisons against unrestricted local-file access. I look closely at individual errors: a plausible finding with the wrong citation needs a different fix from a document the system never examined.

I also check how a reviewer can inspect a finding, correct it, and return to unresolved questions. Building better AI products through evaluation explains how I compare that work with giving an AI model direct access to the same files.

My cofounder and I met as roommates in 2020 and started building together in hackathons before we started Colabra. More than five years later, we've built software used by research labs and private-equity teams. I'm proud of how much we've learned to take on together, from unfamiliar industries to new generations of AI.

We closed an oversubscribed $1.5 million pre-seed round in 2021 without meeting investors in person. Months before the raise, I researched investors in our sector and started building relationships with founders in their portfolios. Those relationships led to roughly eighty warm introductions. I concentrated the investor conversations into three weeks.

Colabra's investors include Village Global, Builders VC, BoxOne Ventures, Inovia Capital, Longtail Health Ventures, and FoundersX. The Wall Street Journal and Forbes covered our first round. In an On Deck interview, I described how we prepared for it.

How we make decisions

My cofounder introduced written Requests for Discussion, or RFDs, for decisions across our remote team. A proposal laid out the decision, context, options, and a suggested approach. Colleagues and advisers could challenge it across time zones. One owner made the call. The final decision went at the top, with the reasoning and disagreements preserved below. New hires could read both what we had decided and why.

What's next for Colabra

Customers increasingly want to build their own AI workflows, using their own templates and interfaces. Our document tools and MCP connection give us a way to support that work alongside our application.

AI is improving quickly, and I think many people still haven't seen how much it could help with their own work. I want to keep improving Colabra and make our tools easier to use inside customers' AI systems. More teams could finish reviews faster, get time back, investigate more questions, and take on work they previously couldn't manage. I'm excited to help them put those capabilities to use.