Agents and pipelines that hold up in production
AI Workflow Development
A demo that works once is not a workflow. I build LLM systems with the boring parts included: retrieval that returns the right chunk, tools with real error handling, evaluation you can trust, and a fallback for every step that can fail.
How I approach it
- Agent design — planning, tool use, memory, and knowing when to stop
- RAG architectures with reranking, chunk strategy and grounded citations
- LangChain / DeepAgents pipelines, plus hand-rolled orchestration when the framework gets in the way
- Human-in-the-loop checkpoints for anything that touches money or customers
- Evaluation harnesses so a prompt change is a measurable decision, not a vibe
Map the decision, not the prompt
Before any model is called I write down what the system decides, what it is allowed to do, and what happens when it is wrong. Most failed AI projects skipped this.
Build the smallest honest version
One path, real data, no orchestration framework. It exposes the retrieval and data-quality problems immediately, which is where the actual work lives.
Instrument before scaling
Traces, token accounting and an eval set from day one. Without them you cannot tell an improvement from a regression.
Harden the edges
Timeouts, retries with backoff, schema validation on every model output, and a deterministic fallback. The model is a component, not the system.
Related work
