Work
What I've Built
Production systems first, experiments second. Every claim here is one I can walk you through.
In Production at AirSprint — built solo
Jetty — Internal AI Agent Platform
AirSprint's internal AI agent platform, built so ~400 employees get AI over company knowledge with zero per-seat licensing cost. One agent with a governed tool belt and contextual query rewriting: document RAG with hybrid retrieval (pgvector + Postgres full-text search, RRF fusion, Cohere rerank), text-to-SQL tools over the data warehouse, an email-gated read-only Salesforce SOQL tool, and an employee directory — a single Python Lambda brain (Docker) on AWS Bedrock, fed by a SharePoint → S3 → pgvector ingestion pipeline. The architecture evolved from a router + specialist agents to a single agent with tools as tool-selection reliability improved, and it's validated by a golden-question regression suite plus stress tests.
AirSprint Data Platform
The company data warehouse, designed and stood up from nothing — a medallion-architecture (Bronze/Silver/Gold) platform on PostgreSQL/RDS consolidating Salesforce, FL3XX flight-ops, and HelpScout, built to retire a fragile Snowflake/dbt pipeline. All infrastructure self-provisioned: in-VPC EC2 ingestion server, S3 data lake, private networking, key-less IAM security (zero static credentials), a cross-compile → S3 → SSM deploy pipeline, and failure alerting. A generic Go CDC pipeline does 15-minute incremental upserts, replacing ~5 legacy full-snapshot jobs with one mechanism. Key metrics reconciled exactly to legacy numbers before cutover.
FlightDeck + API Server
AirSprint's internal operations platform — I'm its sole developer (Go backend + Next.js frontend) — and the OpenAPI-driven Go API server it runs on — the single data path between apps and the warehouse, decoupling serving from ingestion. Shipped features include passport MRZ scanning with eAPIS XML conversion (AWS Textract), an employee empty-leg booking perk, a per-diem/empty-leg payroll engine, bitmask route-level RBAC including a ~1,600-user owner-portal roster, and fleet-utilization and crewing dashboards. Absorbed an external Salesforce contractor's work — replicating Salesforce workflows in-app to sharply cut license spend — with a strangler rewrite of the contractor-built client booking application underway.
Personal Builds & Experiments
PolyEdge — Prediction Market Council
A multi-agent engine for any Polymarket market. Most prediction bots ask one LLM "what's the probability?" — a dressed-up coin flip. PolyEdge instead assembles a panel of experts tailored to each market — a semiconductor analyst, hedge-fund PM and short-seller for a chip market; a Pentagon advisor and IRGC commander for a conflict — each reasoning in character on a different AI model (via Groq), then weights their consensus and compares it live against the market price. Search any market and watch the council deliberate. Deployed on AWS Lambda + API Gateway.
ev-rag — EV Knowledge RAG
A retrieval-augmented bot for electric vehicles — engineering, brand/model differences, charging, and FSD/autonomy. Ask anything and it retrieves from a 44-source, ~2,950-chunk corpus and answers with inline citations, refusing to fill gaps from general knowledge. Auto-routes queries to metadata filters, balances cross-brand comparisons so neither side is starved, and is validated by a golden-set eval. Runs on local BGE embeddings → Supabase pgvector → DeepSeek, served from AWS Lambda.
Moodline — Tennis Market Psychology Engine
An experiment in what betting markets miss: the human. Markets price the statistics; Moodline prices the psychology — deterministic signals (fatigue, marathon matches, head-to-head demons, streaks) plus AI agents reading each player's recent headlines nudge the market's devigged baseline by at most ±6%, never overriding it. Picks lock before first serve into an insert-only public board — losses never deleted — and are graded nightly. Backtested on ~48,000 ATP/WTA matches of real closing odds with the honest result published: the market mostly wins, but where the model disagrees hardest the disagreement carried signal, and the market appears to overprice the hot hand.
The AirSprint systems are internal, so there's no repo to click — but I'm happy to whiteboard any of them, tradeoffs included.
Get in Touch