Builder · AI Products · 0→1

I build AI-driven products end to end — model to interface.

I'm Rafael Leite. I take a hard problem, build the full product around it — trained models, statistical systems, the app and platform, the economics — and ship it to real users. Most recently, an entire chess-training engine running on AI I trained myself.

Full-stack developer & data scientistBrazilian Chess National Master · 2300 chess.comEB-1 extraordinary-ability immigrant
01 Flagship

RivalChess

Solo-built · in productionrivalchess.app ↗

A mobile chess-training app powered by a proprietary AI stack I designed and trained from scratch — not an OpenAI wrapper. Adaptive opponents, move-level coaching, and endgame intervention, now running entirely on-device at zero marginal cost.

0
own-trained human-like models, 600–2400 ELO
$0.00
marginal cost per game — runs fully on-device
0
servers — no backend inference at all
0%
model + product owned, no vendor lock

Human-like opponents

Nineteen models I trained from scratch, one tuned to each rung of the rating ladder — so a 1400 opponent plays like a real 1400, blunders and all. That's the whole point: most apps just dial a single engine down in strength, and it plays alien, inhuman moves no person would ever make. Mine reproduce how players actually think at each level. A custom game-phase predictor and ACPL calibration then decide when and how to coach.

PyTorchhuman-like modelsgame-phase predictorACPL calibration

Human-like move timing

A dedicated model simulates how long a real player would actually think on each move — conditioned on their rating, the time left on the clock, the phase of the game, and the complexity of the position. It's the difference between a bot that feels mechanical and an opponent that feels alive.

ELO-conditionedclock-awareposition complexitygame phase

The cost engineering

I drove inference cost from ~$0.000195 per game all the way to zero. The models now run embedded on the device, paired with a Stockfish WASM build — no backend, no servers, no per-game compute bill at all. A premium product with zero marginal cost is a product nobody can undercut.

on-device inferenceStockfish WASMembedded modelszero backend

The product & platform

Full migration from bare React Native to Expo / EAS for a maintainable release pipeline. Mandatory auth (Sign in with Apple / Google), a 30-game free tier, and monetization wired natively per platform — native in-app purchases on iOS and Android, and Stripe on the web.

React NativeExpo / EASnative IAP (iOS/Android)Stripe (web)

Why it matters

RivalChess is the proof of the thing I do: own the whole vertical — research, ML, infra, app, pricing — and make decisions across all of them at once. One person, the full loop from a trained model to a shipped product.

0→1solo founder-engineerapplied ML
02 Track record

Things I've built and shipped.

Product · Mobile · Live

8Chess

My first chess app — built, launched, and grown into a real user base before RivalChess. The product that proved the playbook I now run end to end.

~55K
users
~200
paying subscribers
See it live · 8chess.app ↗
Audience · Media · Live

Xadrez Brasil

A YouTube channel I run as a parallel business — produced, edited, and grown into one of the largest chess channels in Portuguese. Distribution I built from zero.

~350K
subscribers
~1M
views / month
Watch the channel ↗
Recognition · Immigration

EB-1 extraordinary ability

U.S. permanent residency granted on the extraordinary-ability standard — earned through chess achievement and entrepreneurship. The U.S. government's bar for “top of the field,” cleared.

2300
chess.com rating
EB-1
green card
03 Current role

Claims Analytics Manager — Windward Risk Managers

Since 2025 · in well under a year

The title says "analytics." The reality is broader: I built the company's entire data & engineering capability from scratch — the pipelines, the production models, and a set of custom internal tools that didn't exist before I arrived. One person, end to end, in under a year.

Data pipelines & automation

Designed and maintain the full data backbone — SQL views feeding 12+ operational dashboards across Claims, Legal, Actuarial and Operations, plus Python orchestrators that run the daily ingestion across multiple source systems automatically.

SQL ServerPythonETL automationPower BI

Predictive modeling, in production

Built and deployed an XGBoost claims model that runs live in operations and drives real triage decisions — including the statistical work behind it: leakage fixes, censoring corrections, survival analysis and honest validation, not a notebook demo.

XGBoostsurvival analysismodel deployment

Custom intelligence tools

Built interactive network-graph tools that map the hidden relationships between entities in claims — adjusters, attorneys, firms, contractors — surfacing patterns for fraud and litigation teams. The kind of tool you can't buy off the shelf, so I built it.

force-directed graphsvis.jsnetwork analysis

Tooling & modernization

Replaced brittle legacy workflows with self-contained web tools — migrating an Excel/VBA legal form into a clean web app wired to automated routing, and rebuilding reporting that several teams now depend on daily.

HTML/JS appsPower Automate0→1 internal tooling
04 How I work

A lab, not a ticket queue.

/01

Own the whole problem

I'm at my best owning a product across its full depth — the model, the infrastructure, the interface, the economics — instead of one narrow slice. The leverage is in connecting them.

/02

Run experiments

Build a lab, form a hypothesis, test cheaply, read the signal, iterate. Most of my best decisions — like the per-game cost work — came from treating the product as something to be measured, not assumed.

/03

Ship to reality

A model in a notebook isn't done. I care about the thing reaching real users at a cost and quality that hold up. Production is the only benchmark that counts.

I've never met a problem I couldn't solve — and I don't believe one exists.

That's not bravado. It's how I've worked for years. Every system on this page started as a problem I hadn't solved yet — then became one I had.

05 About

Brazilian-American, in South Florida.

I came to the U.S. from Brazil and earned permanent residency on the extraordinary-ability standard — recognition built on a chess career and on the businesses I created around it. My foundation is engineering: I'm trained as a Production Engineer at Poli-USP — the engineering school of the University of São Paulo, the most selective in Brazil and consistently the top-ranked university in Latin America. I'm also a Brazilian chess National Master (2300 on chess.com), and over the years I became a full-stack developer and data scientist building products at the intersection of all of it.

What runs through all of it is the same instinct: I like hard problems, and I like building the entire machine that solves them. That's what RivalChess is, that's what I built for an insurer's analytics function, and that's the kind of work I want to keep doing — remote, creative, close to the product.

Production Engineering · Poli-USP (top university in Latin America)
Brazilian National Master · 2300 chess.com
EB-1 extraordinary ability
Claims Analytics Manager · Windward Risk Managers
Based in Parkland, FL · remote-ready
06 Contact

Building something hard? Let's talk.