Journey

At 24M, I worked in advanced battery development and served as technical lead on a lithium-metal battery project funded through the U.S. Department of Energy's ARPA-E. The original $3.5 million award brought together 24M, Sepion Technologies, Lawrence Berkeley National Laboratory, and Carnegie Mellon University.

Lithium metal offered a way to store more energy in a battery, but it could form dendrites: branching metal structures that can cause a short circuit. The project focused on a protective layer of polymer and ceramic that would block dendrites and reduce the chemical reactions that degrade the cell.

My work included material evaluation, cell testing, failure analysis, and cost optimization. The protection also had to be inexpensive to manufacture at scale, so the project aimed to use conventional roll-to-roll production processes.

I began to feel that advancing in research would mean going back for a PhD. Instead, I started teaching myself data science and machine learning at night. That eventually led me to a different part of the energy problem: financing projects.

I founded Jumpstart Energy in 2018 to connect clean-energy projects with capital. I found prospective developers and investors through sustainability conferences and solar-industry rankings. Over about seven months, I contacted more than 800 solar project developers and 5,000 enterprises. We built a network of 33 developer partners contributing projects and began conversations with five Fortune 100 companies interested in purchasing clean energy.

But when I asked what it would take to commit capital, I began to understand why introductions weren't enough. Developers could already find financing for the projects investors considered safe. Where developers wanted help was the long tail of projects that nobody wanted to invest in. Investors wanted low-risk projects, lengthy diligence, and established relationships. Adding more people to the marketplace wasn't going to resolve that mismatch.

I narrowed the business to underwriting for U.S. residential solar loans, where there was more data to work with than in commercial projects. I built a model to help solar companies assess default risk and evaluate borrowers who fell below their credit-score cutoffs.

I submitted requests for datasets from government solar programs and built an initial model using credit-report variables, census data, local weather, and national economic data. The model combined gradient-boosted trees and feed-forward neural networks to estimate each borrower's probability of default.

I used that first model and a whitepaper to persuade companies to share their own data. Those datasets let me improve the model and offer an analytics service to solar companies. After about six months, we had four paying enterprise customers. We bootstrapped the business to profitability.

Scaling the business meant working with much longer timelines. Many solar loans hadn't been around long enough to observe their full repayment history, and investors wanted years of performance data before committing capital on the strength of a new model. Proving the model through loans we financed ourselves would have required capital we didn't have.

I sold the company in 2019. We spent several months helping the buyer integrate the underwriting model and plan its future machine-learning work.

After selling Jumpstart, I joined Mission Lane as a senior data scientist, working on fraud and credit decisioning in 2020 and 2021. I chose financial services to deepen my statistical modeling skills alongside experienced practitioners and expand my network beyond energy.

I was the only person on my immediate data-science team without a graduate degree. I built the first-generation fraud model we deployed and a third-generation credit-evaluation model.

In consumer lending, the Equal Credit Opportunity Act restricts how lenders use applicant data and prohibits discrimination in credit decisions. Before adding a variable to a credit model, we considered whether it was appropriate and what consequences it might have for applicants. Our compliance team checked the models for discriminatory effects. When we declined an application, we had to give the applicant specific reasons. We spent longer building models because we also had to understand and explain their decisions. For fraud models, we also had to consider how many flagged applications our reviewers could investigate.

When I left in January 2021 to start my next company, my manager invited me to work with him again. I remember the chief data scientist telling me I was welcome back if the company didn't work out, and suggesting a three- to four-month sabbatical. Mission Lane's head of engineering also reached out after hearing I was leaving and expressed interest in investing in my new company.

My cofounder and I met as roommates at a six-week hacker house in Tulum during the pandemic. We both preferred experimenting with GPT-3 to going to the beach. I wanted to build a company with him even though I still felt I had a lot to learn at Mission Lane.

Within two weeks of the Colabra idea, he had built an MVP and the brand while I spoke with potential customers to test whether the problem mattered to them.

We started Colabra as an AI company focused on scientific R&D. I helped raise our oversubscribed $1.5 million pre-seed round and won customers for the product, including a paid contract with the Yamanaka lab.

Conversations with a firm acquiring R&D software companies introduced us to M&A data rooms. We saw an opportunity to apply our work on documents, search, and permissions to diligence, and raised more money to pursue it. Our team went on to build and deploy software for PE customers that analyzes tens of thousands of financial and legal documents.

I've kept building relationships, shaping the product, and finding new ways to reach customers as the company has evolved. More than five years into building with my cofounder, I'm proud of what we've built at Colabra and excited to keep developing it. The Colabra case study tells that story in more detail.