AI Engineer — CCSL, Federal Institute of Rio Grande do Norte (IFRN)
Public-sector R&D centre. Primary engineer on SmartFAQ, a citizen-facing AI assistant for Brazil’s Ministry of Communications, in production and used by the public.
- Designed and shipped the production RAG pipeline: LangChain and Milvus for retrieval, BGE-M3 embeddings, Llama 3.3 70B served through Ollama, behind Flask services with Celery and Redis queuing and PostgreSQL as the system of record.
- Built the retrieval evaluation harness — context precision and recall over a fixed question set, plus an LLM-as-a-judge with calibration tests measuring the judge’s own effect on the scores.
- Instrumented end-to-end observability across the services in the request path: a propagated request ID, structured trace events with per-phase latency, and a provenance field distinguishing true retrieval from fallbacks and errors.
- Enforced a fail-closed retrieval guarantee, covered by automated conformance tests, so the service declines to answer rather than producing text with no retrieved source behind it.
- Profiled the pipeline end to end and prioritised the optimisation backlog by measured impact rather than intuition.
- Shipped the governance layer — role-based access, soft delete, deduplication hashing, index integrity under concurrent rebuilds — and the ingestion layer: Selenium crawlers over JavaScript-rendered government portals feeding an ETL into the structured knowledge base.
Authored twelve technical reports on the system’s architecture, audits and performance, and co-writing a research paper on the platform.
The system and its source are client-owned and private, so this describes the engineering I was responsible for rather than the implementation.