Summary
I build the software around foundation models: agent harnesses, retrieval-augmented generation (RAG) over your own data, Model Context Protocol (MCP) servers, and the evals that catch a regression before a release does. I build my own tooling the same way, with custom Claude Code skills and agent loops that stop for a person before anything lands. Over six years in fintech taught me to treat the safety checks as part of the feature, not an afterthought.
Experience
Efex Software Engineer · Mexico City Jan 2025 – Present
- Wrote the MCP server that lets an AI assistant query our financial APIs: 12 tools, auth fallback chain, scoped access, circuit breaker.
- Extended cross-border payments: opened the Mexico to China corridor, added COP to USD netting that retired a treasury process converting payments one at a time, and reworked MXN to USD netting so eligible payments clear in about two seconds instead of five minutes.
- Built bulk payments end to end: a Django and Celery pipeline that validates spreadsheet uploads row by row, checks balances and executes partial batches, and the React flow that tracks each upload through validation, execution and reconciliation.
- Authored the team's permissions library, RBAC and ABAC on CASL with a declarative React API, and replaced the dashboard's ad-hoc role checks with it.
Cicada Software Engineer · Mexico City 2023 – 2024
- Bond trading platform: real-time chart visualizations, and an Electron desktop app with event-driven RFQ notifications and zero message loss.
Konfío Software Engineer · Mexico City 2022 – 2023
- Cut product funnel load time 15% and raised conversion 10%.
Clara Software Engineer · Mexico City 2021 – 2022
- Corporate spend management; cut transaction monitoring refresh latency 20%.
What I’m building
- Collaborative notes with an AI copilot, live at knowtis.app: a Tiptap editor with real-time collaboration over CRDTs, a NestJS API, Postgres and Redis, and offline-first editing.
- Bring your own key across OpenAI, Anthropic, Google and OpenRouter, behind a provider fallback chain with per-user token and cost budgets.
- Hybrid retrieval, Postgres full-text search fused with pgvector through Reciprocal Rank Fusion, and nightly Promptfoo evals for grounding, tool choice and prompt-injection resistance.
- Study tools that turn notes into flashcards, quizzes, summaries and mind maps with spaced repetition, and an MCP server that exposes notes to any MCP client, from desktop assistants to IDEs, over OAuth scopes.
How I work
Claude Code daily, with Codex and OpenCode alongside. Custom skills built on top of the superpowers and Anthropic plugin ecosystems drive my own delivery loop: resumable, with approval gates, and a QA harness that trusts an independent oracle over the output of the command that just ran. MCP clients wired to Jira, Playwright, PostHog and Figma.
Skills
- AI & LLM
- RAG (pgvector, Postgres full-text search, Reciprocal Rank Fusion) · agentic tool use · MCP servers and clients · evals (Promptfoo) · prompt-injection defense · multi-provider failover · structured outputs · prompt caching · token and cost budgets · Vercel AI SDK · Anthropic · OpenAI · Google · OpenRouter · FastMCP
- AI tooling
- Claude Code · Codex · OpenCode · Cursor · custom skills and plugins
- Languages
- TypeScript · Python
- Backend
- NestJS · Django, DRF · Celery · PostgreSQL · SQL · Redis · MongoDB · DynamoDB · AWS SQS · WebSockets · Yjs CRDT
- Frontend
- React · Next.js · Vite · TanStack · Zustand · Redux · Tailwind · Tiptap · Electron
- Platform
- AWS (Lambda, Cognito, S3, CloudWatch) · Docker · GitHub Actions · Vercel · Nx · Vitest · Playwright · PostHog flags
Education
Instituto Politécnico Nacional, IPN · Electronics and Telecommunications Engineering · 2009 – 2014
Spanish native · English professional working proficiency