Building in fintech and sports — an AI copilot for commercial credit underwriting (Bashi) and a free-to-play, ranked fantasy app (Snappy) currently in beta.
I studied computer science at UC Berkeley, by way of an International Baccalaureate diploma. I now spend my days writing the code, the pitch decks, and the cold emails for two early-stage companies that have nothing in common on the surface — and almost everything in common underneath.
Bashi turns a multi-week credit-underwriting workflow into an afternoon. Claude-powered extraction, deterministic financial modeling, and an output that looks like the memo a senior credit officer would write.
Snappy is built on a contrarian bet — that the right form-factor for fantasy is short, head-to-head, ranked, and feels like a mobile game, not a spreadsheet. Live matchmaking, snake drafts, real-time scoring.
Both products live where messy real-world data meets a product surface. That's the work I'm interested in — turning the hard, schema-less middle of a problem into something a normal person can use.
An AI copilot for commercial credit underwriting.
“Unlocking potential, together.”
— bashi.app, public homepage. The product itself is intentionally quiet for now.
fig 01 — Bashi extraction pipeline, illustrative
Commercial credit underwriting is a multi-week ritual. Analysts re-key numbers from PDFs into Excel, build 3-statement projections by hand, and write a memo. The work is tedious and error-prone — the slowest part of every middle-market deal.
Bashi reads the PDF the way a credit officer would. Vision-language models pick the right pages. Type-checked extractors lift the numbers. A deterministic engine builds the projection in pure Python. The model writes the memo skeleton — never the math.
We turn a two-week pass into an afternoon. The model does what models are good at — read mess, structure it. The deterministic layer does what models shouldn't — arithmetic, projections, audit trail. The human is freed to do the only part that still matters: judgment.
Fantasy sports, ranked. Pick three real players. Go head-to-head. Climb from Bronze to Legend.
fig 01 — Snappy live-scoring engine. Real formulas, simulated stats.
fig 02 — iOS, current build
Hit Battle. Ranked 1v1 matchmaking at your ELO. Eight divisions from Bronze to Legend. Win and climb, lose and learn.
Snake-draft your roster from real players in live games. Twenty seconds per pick. Every choice matters.
Watch it play out live. Every pitch, every play scores in real time. Your roster against theirs, fantasy points rolling in.
DraftKings won by being a spreadsheet. The next generation of fantasy users won't sit down for a 12-hour slate. They'll play in the queue at Chipotle. The product needs to feel like a mobile game, finish in minutes, and make every play feel consequential — without anyone risking a paycheck.
Snappy ships two formats. NBA quarter-only contests resolve in roughly 30 minutes. MLB head-to-head snake drafts finish in three innings. Both are 1v1. Both have a 20-second draft clock. Both are free to play, with rank — Bronze through Legend — replacing dollars as the thing on the line.
Live matchmaking with a bot fallback at 90 seconds. A Glicko-2 rating that buckets players across eight divisions. Three independent scoring engines. Real-time stat polling against league feeds. Audit-grade tests on the server so the rating math never slips.
Laser tag already exists. We tore apart TENS units, wired the electrode pads through Arduino relays, and built a tag game where getting hit means getting shocked.
fig 01 — Lobby state, simulated. Real games include a real flinch.
Three stacks talking to each other in real time is the most humbling thing you can build — every clock skew, every dropped packet, every flaky IR hit, every misfired relay surfaces through whichever layer is loudest. You learn to instrument before you optimize. You also learn that physical feedback fixes a lot of UX bugs.
Laser tag already exists. The whole game is solved at the toy-store level: photons hit a vest, a buzzer beeps, you keep playing. We wanted stakes that the body actually noticed — something between the photon and the score. The shock is what made it a game we actually wanted to play.
Each player wears an IR receiver wired to a torn-down TENS unit — the over-the-counter electrode-pad device physiotherapists use for muscle stim. The Arduino sits between them. When the IR sensor catches a coded pulse, the Arduino briefly closes a relay and the TENS unit fires through the pads on your forearm. Current-limited and within consumer-device specs — sharp, not dangerous.
A Flask + SocketIO server runs the room: lobbies, teams, hit-points, modes (team-deathmatch and capture-the-flag). Every IR event routes through the server before it can fire any TENS unit, so the rules of the game live in one place — and the body trusts that singularity.
Each player carries an iOS HUD — health, scoreboard, respawn timer. The phone is what you check between rounds. The Arduino is what carries the shot. The pads are what make you remember it. Three stacks, one feedback loop, one stinging forearm.
Hackathon · NL → charts
A natural-language interface for plotting physics. Users typed sentences like "quantum harmonic oscillator" and a fine-tuned Cohere classifier mapped them to chart actions. October 2022 — a month before ChatGPT shipped.
Sports ML · bracket optimization
A full pipeline for the 2026 NCAA bracket. XGBoost win-probabilities calibrated with isotonic regression, blended with Vegas lines, then a 100k-iteration Monte Carlo that picks the bracket maximizing P(top-3 finish) in a 40-person pool.
Hardware · RFID / NFC
A digital vinyl player. Tap an NFC-tagged card on an MSP430 reader and the matching record starts playing — physical artifact for a digital library. Energia firmware, custom RFID handling, hand-cut chipboard "sleeves".
Side project · web game
Wordle, but you're guessing today's UC Berkeley dining-hall menu. Flask, App Engine, a custom on-screen keyboard, and an admin page for the daily answer.
AI · Russian-lit recommender
An OpenAI-backed recommender for niche 19th-century Russian literature. You feed it a list of books you've enjoyed and it returns less-obvious picks in the same vein — and it's hard-coded to refuse to suggest War and Peace, Crime and Punishment, Anna Karenina, or The Brothers Karamazov no matter what you give it.
Research · long-form web
A nine-section research site — country profile, media influence, legal challenges, surveillance, countermeasures — with an interactive Leaflet map and hand-built static pages. The non-code work in the inventory.
Transferred in via Queen's University in 2022; programming-contest history before that.
Three years of submissions across the senior division.
Multi-round competitive programming.
For Bashi, write to adam@bashi.app. For Snappy, adam@snappyfantasy.com. Anything else, the Berkeley address is the easiest to reach.