AI Engineer · Backend · Popayán, Colombia
I put language models into systems you can verify. Since 2022 I have been building backends that turn documents into data for financial and public-sector organisations in Colombia: extraction with LLMs, validation with business rules, and integration with the systems already in place.
Portfolio: 0103juan.github.io/portfolio-site (Spanish and English)
Every number below comes from a real run and is explained, with its limits, in the repository's README.
| Project | What it is | Measured |
|---|---|---|
| consultor-tributario | Spanish RAG over Colombia's tax code, answering with article citations | 29 of 33 questions correct, 100% grounded |
| enterprise-rag | RAG with query rewriting, hybrid retrieval, reranking and claim-level hallucination checks | 87.9% correct, 100% grounded; quality gates in CI |
| secop-analyst | A LangGraph agent over public contracts: hybrid search on pgvector, model-written SQL contained by Postgres, human approval, two-tier memory on Redis and Postgres, OpenTelemetry, local Kubernetes | One real question so far: 38 figures checked against the database, all correct, for about $0.06 |
| model-gateway | The Python package the two RAG projects call models through: routing between a small and a large model, fallback, response cache, cost per call and a budget per key | Same evaluation at $0.66 on the large model and $0.27 on the small one, 28 and 27 of 33 correct; a repeated run cost $0.00 |
| sandboxed-agent | Multi-agent coding system with human approval gates, inside a locked-down Docker sandbox, no agent framework | One full run written up, with its cost |
| legacy-db-mcp | MCP server that lets a model query a legacy database: read-only, column-masked and time-boxed by the database engine | Three real sessions with the model written up |
| secop-mcp | MCP server over Colombia's public procurement data | A real session showed two failures; both were fixed in the tool, not in the prompt |
| verificador-citas | Checks citations of Constitutional Court rulings against the official register, and measures how often a model invents them | Checked against all 29,474 rulings in the register; the model benchmark has not been run yet |
| cifras-verificables | An agent over Colombia's open data where every figure carries a receipt that reproduces it, checked without a model | Tools run against the live portal; not run with a model yet |
Contratos a la vista shows what each Colombian state entity contracts, with whom and how.
- secop-api: Node.js and TypeScript with no runtime dependencies. Live on a free plan, so the first request can take a while.
- secop-dashboard: Angular. Live, tested end to end against the real API on every deployment.
- secop-mobile: Flutter, for Android and the web.
- terraform-cloud-platform: the same platform on Google Cloud and AWS in Terraform (GKE and EKS, Cloud SQL and RDS, Memorystore and ElastiCache, Cloud Run, workload identity, keyless CI), tested against mocked providers. Not applied to an account.
- crediya-microservice-loan: a reactive hexagonal microservice in Java 21 and Spring WebFlux. Approving a loan publishes to SQS and a Lambda emails the client, checked in CI on LocalStack.
- crediya-microservice-auth: users, roles and JWT login, mutation-tested.

