What I'm building
for accelerators
and venture funds.
I've spent twenty years building in hard places, and a couple of years ago the tools changed enough that I could build a different kind of thing. When generative AI got good enough to do real work, I taught myself to ship software with it, and I started making the tools a lean team actually needed instead of waiting for a roadmap that was never coming. The same threshold I caught with LED costs in 2011 and desktop 3D printing in 2017, pointed at software now.
Working builds, designed and built solo.
What's here came out of catching that threshold: working builds and prototypes I designed and built for a global accelerator's real problems, a portfolio of 4,486 startups the team couldn't search, segment, or learn from, with no months to wait. Most are MVPs. They show what I can see, design, and build now that the tools have changed. The record of shipping and scaling in the market lives on the main portfolio.
Three agentic builds.
The tools where AI does the real work: research, scoring, self-critique, orchestration. Each one is something a program team or an analyst used to do by hand.
Crucible
Most accelerator programs evaluate companies the same way: spreadsheets, gut feel, and whoever has the most domain knowledge in the room. Crucible replaces that with something more rigorous. An HC&LS company submits an application; within minutes the system has run live web research on the founders and market, looked up their FDA regulatory pathway, pulled CMS reimbursement data, and scored them across four frameworks: IDEO human-centered design, MIT Disciplined Entrepreneurship, Clinical Science maturity, and a Revenue Bridge analysis. That used to take a program team days.
The architecture is three-pass, not a wrapper. Sonnet with web search runs the initial evaluation and captures every query it ran and what it found. Haiku critiques the output against four specific criteria: FDA pathway accuracy, reimbursement specificity, risk severity calibration, and score-recommendation alignment. If it flags a problem, Sonnet revises only the failed fields. The reasoning trace is visible in the UI, collapsed by default. Specialists, who only see applications the AI flagged for genuine domain uncertainty, see more of it. Their time is treated as the scarcest resource in the system.
The scoring rubric is built around one question: does this company's survival depend on a future binary event it can't control? A regulatory approval it hasn't gotten. A reimbursement code that doesn't exist yet. The top band (90–100) is reserved for companies that don't have that dependency. The demo cohort is 12 real HC&LS companies with publicly disclosed venture raises: Hippocratic AI, Tennr, Abridge, Rad AI, Eko Health, Proprio, Cala Health, Pomelo Care, Nourish, Avenda Health, Proscia, and Boundless Bio. Every evaluation runs on publicly available information only.
Field Notes
A private deal intelligence tool for early-stage VC and advisory work, in private beta. Three capabilities, connected through a persistent company record. First, an investment memo builder structured around the Barley/Roberts HBS framework (why now, why this, why these people, traction read, what would have to be true, risks, recommendation), with inline framework cards citing Andreessen on markets and Botha on what would have to be true, plus an AI-generated IC pitch and skeptical-partner stress test. Second, a term sheet analyzer that puts every clause in plain language and tags it Market, Aggressive, or Founder-friendly against Q1 2026 benchmarks (Aumni, Carta, NVCA), with a VC Lens and Founder Lens toggle and an exit waterfall across four scenarios. Third, a fund profile and portfolio construction model: investable capital, how many companies the math supports, three return scenarios, and a deal-vs-fund read on every term sheet (ownership achieved versus target, percent of fund consumed, reserves needed, the multiple this one deal has to return to carry the whole fund).
The integration is the point. Cap table tools (Carta, Pulley) and term sheet templates (NVCA, YC) exist separately. None connect memo writing to term sheet analysis to fund construction logic. The framework layer is the through-line: the tool encodes established VC memo and term-sheet frameworks, with source attribution visible in the UI. Waterfall math handles the common case correctly and explicitly flags edge cases, like pari passu structures and participation caps, rather than producing confidently wrong numbers.
Built in extended conversation with Claude across a single working session. Built across three iterations: v1 (memos, term sheets, AI synthesis), v2 (lifecycle events, multi-round stacked waterfall, shareable links with viewer tracking), v3 (fund profile, portfolio construction, deal-vs-fund integration, in-app admin invite system). An early scoring prototype was scrapped for being too accelerator-oriented and rebuilt from investor-standard memo structure. About 4,500 lines of TypeScript/TSX.
Accelerate v2: Program Management Platform
MassChallenge runs twelve programs a year across Boston, Texas, Israel, and Switzerland on seven disconnected systems. No single view answers "how is this cohort doing" without opening four tabs. Accelerate v2 pulls the program work into one platform, keeping HubSpot as the CRM and Circle as the alumni layer, and replaces AcceleratorApp, the legacy Accelerate platform, and the Airtable bases that filled the gaps between them. There's a hard deadline: the AcceleratorApp contract ends July 2026, after HC Traction 2026 closes. HC Traction was the only program ever run on it, so there's no migration to manage. After that, everything runs on v2.
It started as a single-file, click-through prototype and is now live as an invite-only functional sandbox: real sign-in and roles, shared data, agents calling Claude through a Netlify AI Gateway with no keys in the browser. Seven personas, each with a role-scoped view, from the applicant wizard and judge queue to the program-manager console, mentor view, founder view, scoped partner portal, and admin console. Three program shapes, each a template that pre-configures modules, agents, and defaults: Traction (20–40 startups, four months, mentor-centric), Challenge (multiple corporate partners, each selecting a handful of startups, three months), and Custom (single partner, built to launch in under two weeks from contract).
Eight AI agents sit underneath, each surfacing a recommendation a human accepts, edits, or rejects. Nothing meaningful runs without a sign-off. They target the single largest documented cost in running a program: the 8 to 10 hours a week staff spend moving and reconciling data between systems, with a target under two. Three are load-bearing and built first. Screener triages incoming applications, aiming for 85%+ agreement with PM decisions. Matchmaker pairs mentors to startups with a rationale on each, targeting 70%+ accepted as-is. Coach drafts the biweekly progress report from session activity so a mentor edits rather than writes from scratch. Allocator, Synthesizer, Curator, Scout, and Conductor handle judge load, score roll-ups, partner fit scores, the alumni-to-mentor pipeline, and orchestration. Agent output is the only thing rendered in MC Purple, so it's always clear what's machine-generated.
I built the prototype and ran the requirements work behind it. It's now the basis for the production version, which I'm driving with the tech team.
One dataset, three tools, built in order.
The accelerator had 4,486 companies in a database with 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.
AI-Powered Startup Sector Classifier
Designed and ran an LLM classification pipeline across 4,486 startups: 5 strategic sectors and 5 healthcare sub-sectors. The pipeline feeds elevator pitches and Crunchbase descriptions to Claude Sonnet in batches of 30–40, with structured output parsing, retry logic, and progress tracking. It replaced a keyword-matching approach that had a 26% error rate on reclassifications. That 26% made the database effectively unusable for any segmented analysis.
This is the foundation layer. The Alumni Explorer, the Portfolio Intelligence Dashboard, and four published LinkedIn articles all run on the dataset it produced. Get this wrong and everything downstream inherits the error, which is why it got built first.
MC Alumni Explorer
Type "spinal cord injuries" and get back 13 companies grouped by relevance tier: high, medium, loosely related. Each card shows one sentence of AI reasoning for why it matched. That's the semantic search layer: natural language queries against 4,486 companies, relevance scored and explained. Staff who don't know the sector taxonomy find what they need without knowing what to call it. Staff who do know it get keyword search, eight filter dimensions (sector, sub-sector, funding stage, geography, demographics, confidence score, exit status, program year), card and table views, starred companies, and CSV export alongside.
Each company card shows funding, last funded date, Crunchbase rank, and a survival confidence score from a cohort-adjusted composite model. A button in each detail modal triggers a Claude API call with web search tool use to find and summarize recent press coverage for that company, on demand. One search, one company, live results.
Portfolio Intelligence Dashboard
A 6-tab interactive dashboard analyzing 15 years of startup outcomes: 4,486 companies, $27.1B in funding, 21 interactive charts. Tabs cover portfolio growth, challenge areas, funding pipeline, survival benchmarks, demographics, and cohort composition. Built as a public-facing data product for VC, corporate innovation, and startup community audiences, and as the analytical foundation for four published LinkedIn articles on portfolio data.
The analytical work underneath: a cohort-adjusted survival model calibrated against BLS and Stripe benchmarks (explaining why Crunchbase's raw 87% active rate is actually closer to 59%), a funding pipeline funnel (4,486 companies, 300 Series A+, 12 IPOs), a power law concentration curve (top 1% holds 45% of portfolio funding), and a funding parity analysis by gender and race across all 5 challenge areas.
Tools I actually used.
No scaffolding for the sake of it. Everything here is a deliberate choice for 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 the whole database unusable. Crucible went from idea to a working system in about seven sessions. None of this was a technical problem at heart. The problems were operational, the kind I've been solving for twenty years, and AI collapsed the distance between seeing one and fixing it. That's what I'd bring to a fund or an accelerator: the speed and reach of one person who can now build the tools instead of writing a spec and waiting.