Senior Full-Stack / AI Engineer — Production Ad-Ops SaaS (React + Supabase + LLM agents)

The product A multi-tenant platform that runs paid social for DTC brands end to end: generating the creative, launching it, tracking performance, and adjusting spend. Live in production against real ad accounts. Three things make it more interesting than a CRUD app: An agent that acts, not just advises. It reads live account state, proposes budget and status changes with its reasoning attached, waits for human approval, executes against the ad platform's API, then verifies the change landed and flags drift. Brand-locked creative generation. Every render is pinned to a brand kit, with categorized reference images that each carry different instructions to the model, and a lint gate that refuses to spend a render on an underspecified brief. A feedback loop. Human verdicts plus actual spend and return feed back into which creative patterns work, so output improves for a specific brand rather than drifting generic. Stack React + TypeScript + Vite + Tailwind + Redux Toolkit. Supabase (Postgres, row-level security, Deno edge functions, scheduled jobs). Meta Marketing API. LLM APIs for generation and agent reasoning. What we need Past prototype, into the part that decides whether it scales. You don't need all of these: CI/CD and release engineering — automated function deploys, preview environments, migration discipline. Currently our biggest bottleneck. Self-serve integrations layer — per-tenant connection management with health and onboarding. Evaluation infrastructure — versioned prompts, an eval harness, A/B on variants against real outcome data. Who we're looking for You've operated a multi-tenant SaaS in production, not just built one. You think about what happens on the second delivery of the same webhook, and you've had to reason about what a client can do with credentials that ship in a browser bundle. Bonus if you've worked against the Meta Marketing API — its rate limits, error taxonomy and review process shape a lot of our architecture. Engagement Paid scoped project (1–2 weeks) first, then retainer or part-time if it works. Clear backlog, written architecture docs, inline commentary explaining why. Reply with a system you've operated in production and one thing that broke in a way you didn't expect.