AI agents that ship to production.

Section 01

What I build

Stack

Anthropic SDK, Python (FastAPI) or TypeScript, Supabase, Postgres, Stripe, Sentry. React or Next.js on top when an operator dashboard is in scope.

"Our agent works in the demo. In production it is flaky and we do not know why. I want someone who has shipped this layer before."

Six pieces. Not all of them apply to every project; the discovery call decides which ones do. What I will not ship is an agent system that punts on output validation, treats Stripe webhooks as fire-and-forget, or leaves observability for "later."

• Multi-agent pipelines

  Orchestrator, chain runner with goal propagation, per-client memory, append-only audit log, tool-use loop (web search, web fetch, site crawl). The shape that turns a roster of prompts into a coordinated system.

• Output QC for LLM systems

  JSON schema validation plus an LLM-as-judge with retry-and-feedback. The reliability layer that turns a flaky prototype into something you can put in front of customers.

• Production backend underneath

  Supabase, Postgres, RLS for tenancy, Stripe Checkout with a signed and idempotent webhook ledger, edge functions for the business logic that cannot live in the database.

• Observability and bounded cost

  Sentry on hot paths, structured logging, daily token-budget guard, audit-logged state changes. Failures visible, costs capped, retries explicit.

• EU compliance when relevant

  GDPR Article 17 erasure and Article 20 portability endpoints, retention and orphan cleanup jobs, audit-loggable deletion. Built in, not bolted on.

• Handover

  Migrations runbook, .env template, walkthrough. You can deploy, run migrations, invoke each agent, verify Stripe end-to-end without my hand-holding.