AI engineering for production: agents, RAG, and LangGraph
Remote AI engineer for US companies. I design the loop, wire retrieval, and ship an agent your team can actually run - not a notebook demo.
What this is for
You have a product, a pile of internal knowledge, and a team that is tired of stitching prompts. You need an engineer who has already put agents in production - recon, database chat, support-style RAG - and who will own the backend, not just the prompt.
I work remotely with US hours overlap. You get Slack-ready updates and a repo you can keep, not a black-box agency drop.
How I build agents
The model is the narrator, not the product. Tools do the work, the graph owns state, and the UI streams what happened. ReconX is this pattern in the wild: the agent picks a recon tool, runs it, and explains the finding without hiding the command.
RAG is the same idea with documents. Index what changes weekly, retrieve with filters, cite the chunk, and refuse to answer when retrieval is empty. Fine-tuning comes later, and only if behaviour - not facts - is the gap.
What you get
- Agent loops that call tools, show their work, and keep a human in control
- RAG over your docs, tickets, or database - with source citations, not guesses
- LangGraph (or equivalent) state machines for branching, retries, and evals
- FastAPI or Next.js APIs so the agent lives inside your product, not a chat toy
- Evaluation harness on real queries before you call it done
How we work
01
Scope
We pick one workflow that is costing you time. I map tools, data sources, and what “done” looks like.
02
Build
I ship a thin vertical: retrieval, the graph, the UI. You can click it in the first milestone.
03
Harden
Evals, logging, and a human-approve step where it matters. Then we extend the same loop.
