Guillermo Hoyo Bravo

Sole AI engineer at IDEA: designed, built, and leads the company’s production RAG platform.

• Designed, built, and now lead a production RAG platform on Docker, Azure CI/CD, PostgreSQL, and Qdrant, with an endpoint factory and reusable ingestion/retrieval libraries so new assistants ship from shared components.
• Integrated a RAGAS evaluation module and Langfuse observability so retrieval and answer quality are measured, not assumed.
• Shipped two assistants to production — an HR assistant and an electrical-engineering-rules assistant — with strong measured results [PLACEHOLDER — metrics to confirm].
• Mentor an engineer building a LangGraph agent on the platform.

Joined to lead one AI feature — a RAG chatbot for a multi-tenant payments SaaS. Built and validated; scheduled to deploy to production in [PLACEHOLDER — target window, ~a few months]. Ongoing collaboration, role still being defined (currently unpaid).

• Designed and led the build of the RAG chatbot, AI-assisted and reviewed commit by commit.
• Built hybrid retrieval: dense pgvector plus Spanish Postgres full-text search, fused with Reciprocal Rank Fusion, then a cross-encoder reranker and a relevance gate.
• Designed privacy-by-design: PII scrubbing before the LLM, and data queries executed locally so the external LLM never sees tenant data — a boundary that holds by construction.
• Built a custom evaluation suite — a 150-question golden set, a retrieval probe, and a RAGAS-lite scorer — and an LLM benchmark that showed with data the bottleneck was retrieval (synonyms), not the generation model.

• Extended a production multi-agent system on Semantic Kernel; added plugins for Text-to-SQL (Databricks), document extraction (Azure Blob Storage), translation, SharePoint lookup, and a Service Desk SOAP service in under four weeks.
• Built a Text-to-SQL agent with an enriched metadata layer and a custom evaluation framework with a golden dataset — 100% accuracy on controlled tests before every deploy. Implemented RBAC impersonation so every query runs under the requesting user’s identity.
• Built FastAPI endpoints with async token streaming, Pydantic-validated I/O, selective per-plugin memory (MongoDB), and end-to-end impersonation; deployed via Azure DevOps CI/CD to AKS.
• Contributed to a hybrid RAG pipeline on Azure AI Search (semantic + BM25 + cross-encoder reranking) with permission-aware retrieval; reviewed intern PRs and set plugin contribution criteria.

• Automated data cleaning and processing for large datasets with Python and KNIME; researched OCR digitisation (Pytesseract, OpenCV, Azure AI Studio).