AI Engineer who ships the whole system: agents, RAG, evals, and the economics to run them.

I'm Nicolás Mastakas (Nico), an AI Engineer. I build the system around the model, not just the prompts: agents that do real autonomous work, retrieval that stays grounded, evals that gate quality, fine-tuning when it pays, and the token economics to run all of it in production. I come at it as a product engineer, not a feature engineer: I stay with how the thing actually feels (the timing, the edge case, the line nobody reads), because that's respect for whoever uses what I ship. Six-plus years shipping solo end to end, self-taught in a field that moves every week.

I decide where deterministic code ends and a model begins, then I make the model accountable: measured with evals, bounded on cost, and kept honest with a human in the loop where trust matters.

• Monorepos, one type system: mobile, web and backend share types, so a contract changes in exactly one place.
• Deterministic gates: typecheck, lint, test, build. Non-negotiable, before anything ships.
• Observability from day one: error tables and honest UI states, not a black-box 'unknown' bolted on later.
• Integrations that survive the real world: external syncs treat every webhook as a ping to go refetch, so a retried or out-of-order event converges instead of double-applying (LedgerBridge).
• Provider-agnostic AI: free and local by default, frontier when it earns it, with the trade-off measured (shoebox: ~90% of cloud at $0).
• Agents I can trust: bounded tool-use loops with hard cost caps, untrusted input spotlighted as data, and an eval that gates the PR (rubric: κ 0.81). Not vibes.
• Humans in the loop by design: the model surfaces probability; a person decides where trust actually matters.