I build the AI that makes decisions, and the payment rails underneath it.

A local-work marketplace. People post jobs, nearby workers bid, an AI prices the work and mediates the negotiation. Every layer is mine — frontend, backend, payments, AI, security, infrastructure, and the LLC.

Pricing engine

Resolves an unstructured request into a scoped, priced job. LLM reasoning over live web-search rate data and vision analysis of user photos. Two layers running in production.

Bid mediator

When employer and worker are apart on price, an LLM negotiates between them toward a settlement. The hardest thing in the build and the reason the product exists.

Money movement

Stripe end to end — escrow held on acceptance, released on completion, paid out to connected accounts. Webhook handling and failure states throughout.

Trust layer

ID verification gating posting and bidding. Two-sided profiles, ratings, and reliability scoring on both workers and employers.

Database security

Row-Level Security policies authored across every table in Postgres. Took the security audit from ten findings down to one.

Auth

Google OAuth and magic-link sign-in, with a dual-mode account model — one user, two products, separate permissions and data access per mode.

Trained and evaluated AI music generation models on an ML research team. I built and curated the audio datasets used for training and evaluation, and wrote the structured output-quality assessments that fed back into the models.

Why me on that team

Ten years producing music. Evaluating a generative audio model is a listening problem before it is a metrics problem — I could hear what the model got wrong and write down why in terms the engineers could act on.

What transferred

Dataset curation, rubric design, and structured model evaluation — the same discipline I now apply to grading the outputs of the pricing and mediation systems I build.