Backend software developer. I started in warehouse operations and learned to code by replacing slow, error-prone work with software people use every day. Mostly TypeScript/Node.js, Python, PostgreSQL and Redis.
My work includes the sign-in, payments, and usage-credit system behind an AI product (client work, hired twice), warehouse labeling software I built in July 2025 that the floor uses daily (144,000+ labels since tracking began in August 2026), and Rook, where code checks an AI agent against real game state.
Open to software engineering roles, remote or in Miami, and to fixed-scope contract work. Six client reviews at claygeo.dev.
Email · Portfolio · LinkedIn · X
Payments and usage credits for an AI product: client work, two contracts
Clerk sign-in, Stripe Checkout, and a usage-credit ledger for a product that runs long, expensive generation jobs. A webhook delivered twice cannot grant credits twice, two requests cannot both spend the last credit, and a run that fails to deliver never costs the user a credit: a credit is held when a run starts, settled when the deliverable exists, and released by an inactivity timer if the run is abandoned. The client read the ledger code and ran the test suite before releasing each milestone, then hired me a second time to take the product to production: nginx with auto-renewed TLS, the app under systemd, a managed Postgres that is not reachable from the internet with separate owner and runtime roles, and production Stripe with the key's permissions cut to what the app uses. Both contracts ran entirely in writing.
Warehouse labeling system: case study, code private
Browser software that turns inconsistent CSV and Excel exports into scannable Zebra thermal labels. Operators scan a barcode, resolve the product, assign a batch, and print. Built in July 2025 and in daily use across two facilities; 144,000+ labels since print tracking began in August 2026, about 2,400 on a working day. The case study covers what broke, including a fix of mine that was wrong.
Rook: case study, code private
A model proposes what to do next. Code with no model in it decides whether the proposal can run, carries it out, and checks the result against game memory instead of the model's own account. It rejected 1,003 of 3,500 proposals: about half were the model's errors, about half were gaps in my own checks.
MCP server for Distru's public API. Fixture mode, dry-run writes bound to a confirmation token, JSONL audit logging, and a PO matcher.
An agent for authorized smart-contract security research on local forks. It runs in a sandbox with an iteration budget, writes Solidity, runs forge test, reads the errors, and revises.
A Rust service that subscribes to live Solana swaps on two exchanges over a WebSocket, parses them on a bounded worker pool so a busy slot slows the parsers and never the socket reader, and flags front-run, victim, back-run triples within 3 slots. Idempotent Postgres writes, a Server-Sent Events feed, and a backtest binary with a precision and recall gate that must pass before it goes live.
- Payments, webhooks, and ledgers that reconcile
- Backend services and REST APIs
- PostgreSQL data modeling and transactional workflows
- Queues, workers, retries, and failure recovery
- Checking AI agent output with deterministic code
- Real-time WebSocket systems
- Messy third-party integrations and browser automation
- CI, observability, and operational handoff
Core stack: TypeScript · Node.js · Python · PostgreSQL · Redis · Stripe · BullMQ · WebSockets · Supabase · Docker · Playwright