Systems that let AI agents act safely in production.
Senior engineer on Reddit's Ads API platform. CTO of Frakkr. Ex-OpenAI. I build the MCP servers, the OAuth and identity layer, and the migrations that cut over at 100%.
An agent asks. OAuth decides. The API only hears from the ones with consent.
Built, shipped and operated.
Scroll to move the rows.Swipe through the cases. Open any card for the full case: what it was, what I did, the numbers.
Where identity, APIs and reliability meet.
That intersection is where agents need the most care, and it's where most of my work has been.
MCP servers, tool-use and evals that hold up with real users.
Scopes, consent and machine identities agents can't bypass.
Backfills, flags and a cutover plan that reaches 100%.
OpenAPI contracts, gateway validation, standards partners adopt.
SLOs, burn-rate alerts and rollouts that roll back on their own.
Training and review guardrails so speed doesn't cost correctness.
From one job to three systems at once.
Web apps for multinationals, then an LLM chatbot shipped in a month, then OpenAI's model-quality work. Since 2024, Reddit's Ads API platform. In parallel, CTO of Frakkr and tech lead of Humbral.
Spanish, English, Portuguese · Certified Software Architect · B.S. Computer Science, University of the People
Full experience
Go and Python services, public API design, identity and OAuth, data migrations, gateway and rollout infrastructure, observability. 278 PRs authored, 128 reviewed.
AI automation agency. Agents and automations for companies in Colombia, delivery quality, architecture and the education line.
Athlete hub platform. Core modules, payments integration that cut per-transaction cost by 60%, API latency cut to about a tenth, and the MCP server.
Coaching senior engineers placed at eBay, Reddit and others on AI in their projects; built the internal AI education program, zero to LLMOps in about three weeks.
AI assistant for lawyers. Response capacity and throughput raised roughly ninefold.
Evaluation, testing and feedback workflows for LLM behavior.
LLM application on the OpenAI API; production chatbot shipped in the first month.
Web applications for multinational clients, an LMS built end to end, cloud cost and query performance.
Building something agents will touch?
Tell me what it is and where the risk sits. I'll answer with how I'd design the guardrails, and what I'd ship first.
