RAG systems that still answer correctly after launch.
Most RAG demos work on ten documents and fall apart on ten thousand. I build retrieval-augmented generation systems for startups that need answers grounded in their own data, measured with evals, and cheap enough to run at scale.
Why RAG apps fail in production
Bad chunking that splits the answer across pieces, vector-only search that misses exact terms, no reranking, no evals so regressions ship silently, and context windows stuffed with irrelevant text that raise cost and lower accuracy. Each one is fixable with engineering, not a bigger model.
How I build RAG
Ingestion that respects document structure, hybrid keyword and vector retrieval, reranking, citations back to the source, and an eval set built from your real questions so every change is measured. Then monitoring for retrieval quality, latency and token spend once real users arrive.
Fixing an existing RAG system
If you already have RAG in production and answers are wrong, slow or expensive, I start with an audit of retrieval hit rate, latency and token spend per query, then fix the biggest problems first.
What I deliver
- Document ingestion and chunking pipeline
- Hybrid search, reranking and citations
- Eval set and regression testing
- Context compression and token cost control
- Monitoring for retrieval quality and latency
- Audit and repair of existing RAG apps
Common questions
What is a RAG developer?
A RAG developer builds systems where an LLM answers using your own documents and data, retrieved at query time, instead of only what the model memorised. The work is mostly retrieval engineering: ingestion, search, ranking and evaluation.
RAG or fine-tuning?
RAG is usually the better first step when answers must reflect your current data and cite sources. Fine-tuning helps with style, format or narrow tasks. Many production systems use RAG first and fine-tune later only if evals show a gap.
Which vector database do you use?
Whatever fits your stack and scale. Postgres with pgvector covers many startups; dedicated vector databases make sense at higher volume. The retrieval design matters more than the database brand.
Shipped work
- Grace AI
Memory and RAG design for a contextual AI companion across millions of context turns.
- Socella
Founding AI engineer. Vector search, agentic automation, RBAC and Stripe. $15K+ early revenue, $50K+ raised.
- ScriptAssist
Cut API p95 latency 35% and MTTR 40% on a production platform with LLM pipelines.