Agentic Systems
& RAG
I build agents that do real work — using tools, searching your data and citing their sources — and keep them accurate and grounded, even running fully offline.
Agents that stay grounded and honest
A model on its own hallucinates. An agent wired to retrieval, tools and your actual data doesn't have to. I build tool-using agents and RAG pipelines where every answer is grounded in real sources — with citations you can click through and verify.
These systems are designed to run on your own infrastructure: local embeddings and vector search, local models, and guardrails that keep the agent honest rather than confidently wrong.
What's included
Stack
Local embeddings · vector search · RAG · tool / function calling · llama.cpp · Python · TypeScript · evaluation harnesses
Frequently Asked Questions
How do you stop the agent hallucinating?
Ground it. Answers are built from retrieved sources with citations, and grounding checks reject responses that aren't supported by the retrieved context — the agent is designed to say 'I don't know' rather than invent.
Can it run offline?
Yes. Local embedding models plus a locally served LLM mean the whole retrieval-and-generation loop can run with no external calls.
What can the agent actually do?
Search your knowledge base, call internal tools and APIs, run code, and chain steps to complete a task — scoped and permissioned so it only does what you allow.
How do you measure quality?
With task-based evaluation on held-out examples — checking both answer accuracy and whether citations genuinely support the claims.
