I turn legacy systems into modern platforms and build AI products from the ground up.

Finance social media is full of confident stock picks and no accountability — nobody tracks whether the loudest accounts are actually right. I built a product that grades stock-picking accounts on their realized returns, turning “trust me” into a measurable track record. I owned it end to end: defined the problem, scoped the build, and shipped a live product that ingests influencer posts, classifies each real call with an LLM pipeline, and scores it against the market. Matching each call to a real security is a retrieval problem — a resolver pulls candidate listings from a security master and holds the model to them — and after a bad fuzzy match produced a confident wrong verdict, I gated auto-apply on retrieval confidence. The hardest product decision was discipline over flash — it withholds a verdict until the data clears a real statistical bar, so users get an honest signal instead of false confidence. Today it tracks 225 accounts and has graded ~343K calls across 7,600+ global securities, scoring every call against its sector benchmark rather than raw return, with empirical-Bayes shrinkage and false-discovery control so a few lucky calls can't top the ranking — all under 4,000+ automated tests. Built and run solo, which meant owning every call from data integrity to what we don't ship.

I run a small lawn care business, and I replaced the SaaS CRM it ran on with an operations app I designed and shipped with AI coding agents. As the product owner I defined every business rule: program pricing tiers, per-visit product recipes, scheduling, payment policies, and a field "truck mode" that walks each stop with checklists, tank-mix math, on-my-way texts, and payment collection. It went live in September 2026, migrating 109 invoices, and now runs Stripe card and ACH payments with autopay, two-way SMS, automated reminders, and state pesticide-application records that fill themselves in, with live weather at the property, when a job is marked done. The AI pieces are built as drafts to verify, never as the authority: a retrieval-augmented assistant that answers from pesticide labels and NC State references with numbered citations, scored against a 25-question known-answer eval, plus vision models that read receipts and product labels.

Built an automated system that takes real positions in prediction markets on its own, and owned it end to end. The core product principle was discipline over speed: every strategy runs through a risk-and-order gate and has to prove itself shadow-to-live against real market data before it can risk a dollar. That discipline is literal — backtests price every fill on the executable ask from recorded order books, signals must beat a shuffled null and hold up walk-forward, and nothing risks capital until it has survived weeks of live shadow fills; rebuilding the price layer alone re-scored 133 "winning" strategies to zero. It runs ~189K priced markets across six assets, four languages, and 4,300+ automated tests. Under the hood, a polyglot low-latency stack — Rust execution, Go services, a TypeScript data bus, Python research — where I discover quantitative signals and train ML models on market-microstructure features, then promote them across cloud and local infrastructure.

A job hunt is really a search-and-evaluation problem: too many postings, most irrelevant, each needing a tailored application. I built a system that turns it into rated, ranked, auto-tailored output, pulling from 13+ job APIs and ATS platforms, using an LLM to score every posting against a candidate profile, run a fit analysis, and generate tailored resumes and cover letters, all on a Kanban board. The generator is retrieval-augmented — it embeds a personal career corpus and pulls the most relevant achievements for each posting, so every document stays grounded in real experience instead of a generic template. Built with Claude Code.