What I'm building
for accelerators
and venture funds.
Working tools, most of them live, all built solo. The problems came out of real operating work: a portfolio of 4,486 companies with no reliable way to search or segment it, an evaluation process running on spreadsheets and gut feel, a public federal dataset that is free and close to unusable. Each tool here replaces something an analyst or a program team used to do by hand. The AI layer is additive in every one of them, and each still works with it switched off.
Three builds of my own.
Tools built outside any job description, for the work itself: finding companies, evaluating them, and writing up the decision. Each one replaces something an analyst or a program team used to do by hand.
SBIR Explorer
6,446 US companies holding federal SBIR/STTR awards in health, disability, and education. 17,736 awards, $13.77B, 2015 to 2025, sourced from sbir.gov. For sourcing, the useful question is not who got funded. It is who has third-party technical validation, no dilution yet, and a cliff coming. 780 contracts end in the next twelve months, which is when those companies are most likely raising. That filter does not exist on sbir.gov, and neither does isolating the 1,805 companies with a named university research partner or the 1,044 that are women-owned.
Search a thesis in plain language and every result shows the sentence from the grant abstract that matched, terms marked inline, so relevance is a two-second read instead of a PDF. 48 to 50 of every 50 results carry a real matched sentence. Retrieval is BM25 first and AI second: Haiku expands the query into editable chips, a precomputed static index of 91,014 terms scores in the browser for free, and Sonnet optionally reranks the top 50. Drop a chip and results rescore. Your terms weight 1.0, the model's 0.6. If the budget runs out or the network fails, it falls back to plain BM25 with the matched sentence intact, and that path is tested rather than assumed.
Every dollar figure is federal research funding, never a round or a valuation, stated on each record and carried into the shortlist export. Derived fields say so: NIH institute labels come from NIH for 55% of companies and from a classifier for the rest, marked predicted with confidence shown. A second mode serves patients, families, and clinicians off the same dataset, and it states plainly that a research award does not mean there is a trial to join. Pre-launch auditing caught 2,862 malformed URLs, 875 records listing "NAN" as the funding agency, and 3 companies still labeled active a day after their contract expired. Fixed in the data, not patched at runtime.
Crucible
Most programs evaluate companies with spreadsheets, gut feel, and whoever in the room knows the sector. Crucible replaces that. A company submits, and within minutes the system has run live web research on the founders and market, identified the FDA regulatory pathway, pulled CMS reimbursement data citing an actual CPT or HCPCS code, and scored the company across four frameworks: human-centered design, MIT Disciplined Entrepreneurship, clinical science maturity, and a revenue bridge. That used to take a program team days.
Three passes, not a wrapper. Sonnet with web search runs the evaluation and logs every query it ran. Haiku critiques the output on FDA pathway accuracy, reimbursement specificity, risk calibration, and score alignment. If it flags a problem, Sonnet revises only the failed fields. The full trace is in the UI, collapsed by default. Specialists see only what the AI flagged for genuine uncertainty, since their time is the scarcest thing in the process.
Scoring turns on one question: does survival depend on a binary event the company cannot control, an approval it does not have or a reimbursement code that does not exist yet. The top band is reserved for companies without that dependency. The demo runs on 12 real HC&LS companies, all publicly raised Series A, evaluated on public information only.
Field Notes
A private deal intelligence tool for early-stage investing and advisory work, built around one persistent company record. A memo builder on the Barley/Roberts HBS structure (why now, why this, why these people, traction, what would have to be true, risks, recommendation), with an AI-drafted IC pitch and a skeptical-partner stress test. A term sheet analyzer that puts every clause in plain language and marks it Market, Aggressive, or Founder-friendly against Q1 2026 benchmarks from Aumni, Carta, and NVCA, with a four-scenario exit waterfall and a founder/investor lens toggle.
The fund layer is what ties it together. Every term sheet gets read back against portfolio construction: ownership achieved versus target, percent of fund consumed, reserves needed, and the multiple this one deal has to return to carry the whole fund. Cap table tools and term sheet templates exist separately, and none of them connect memo writing to terms to fund math, which is where the decision actually gets made. The waterfall handles the common case and flags edge cases like participation caps and pari passu structures instead of returning confidently wrong numbers.
One dataset, three tools, built in order.
An accelerator with 4,486 companies in a database and no reliable way to search, segment, or learn from it. Fixing that was one connected piece of work, and the order mattered. The classifier had to be right before anything built on top of it was worth building.
Startup Sector Classifier
An LLM classification pipeline across 4,486 startups: 5 strategic sectors and 5 healthcare sub-sectors. Elevator pitches and Crunchbase descriptions go to Claude Sonnet in batches of 30 to 40, with structured output parsing, retry logic, and progress tracking. It replaced keyword matching that carried a 26% error rate, which had made the database unusable for any segmented analysis. Everything downstream, the Explorer, the dashboard, and four published articles, runs on the dataset it produced, so it got built first.
Alumni Explorer
Type "spinal cord injuries" and get back 13 companies grouped by relevance tier, each card carrying one sentence of AI reasoning for why it matched. Staff who don't know the taxonomy find what they need without knowing what to call it. Staff who do get keyword search across eight filter dimensions, card and table views, starred companies, and CSV export. Each company shows funding, last funded date, Crunchbase rank, and a survival confidence score from a cohort-adjusted model. One button per company triggers a live Claude call with web search to summarize recent press on demand.
Portfolio Intelligence Dashboard
Six tabs and 21 interactive charts over 15 years of startup outcomes: 4,486 companies, $27.1B in funding, covering growth, sector mix, funding pipeline, survival, and demographics. The analysis underneath is the part that matters. A cohort-adjusted survival model calibrated against BLS and Stripe benchmarks put the real active rate near 59% against a raw Crunchbase read of 87%, and that lower number is the one that got used. Alongside it, a funding funnel (4,486 companies, 300 Series A+, 12 IPOs), a power law curve showing the top 1% holding 45% of portfolio funding, and a funding parity analysis by gender and race.
Tools I actually used.
Every choice here answers a specific constraint: single-file deploys, no build pipeline, serverless for anything that touches an API key.
What I'd bring to a portfolio.
Being able to build the fix instead of commissioning it changes what a lean team can attempt.
The Explorer took a week and replaced a workflow that would have taken a quarter to spec and procure. The classifier took two days and fixed an error rate that had made a whole database unusable. Crucible went from idea to working system in about seven sessions. None of these were technical problems at heart. They were operational, the kind I have been solving for twenty years, and AI collapsed the distance between seeing one and fixing it.