Multi-agent AI systems that hold up in production.
I design and build LLM agent systems — orchestration, retrieval, shared memory and critic loops — and ship them to real users. The engine behind PromptMotion, my own product, runs a team of specialised agents that plan, question and review each other's work — launched in 5 months, 1,500+ users in its first 2 months.
- 1,500+
- users in 2 months on a multi-agent product I built
- 5–10
- specialist agents per run in one production engine
- 6+ yrs
- building AI & computer vision
Problems I'm usually brought in to solve.
One big prompt isn't enough
Your LLM feature works in demos but output is shallow or inconsistent at scale.
Work spans many steps
The task needs planning, tool calls, retrieval and state that carries across steps.
Quality is hard to control
You need outputs checked for correctness — and for whether they're actually good.
Costs grow with usage
More agents means more LLM calls, and your cost per run needs to come down.
What you get.
01
Agent architecture & orchestration
Specialised agents with narrow jobs, orchestrated with LangChain and LangGraph.
02
RAG knowledge bases
Domain knowledge agents retrieve from, instead of stuffing it into prompts.
03
Shared state & memory
An append-only ledger every agent reads, so output stays consistent across many steps.
04
Critic & evaluation loops
Separate critics for correctness and for quality that send work back until it passes.
05
Cost & latency engineering
Caching and model routing so a multi-agent system stays affordable at scale.
06
Production delivery
APIs, monitoring and the product around the agents, not just a notebook.
01
Discovery callWe talk through the problem, your data and constraints, and what success looks like. You get an honest view of what AI can and can't do here.
02
Architecture & planA clear design — components, models, data flow and how quality will be measured — with a scoped first version.
03
Build & evaluateWorking software early, tested against real examples from your domain rather than demos.
04
Ship & improveDeployed to production, monitored, and tuned for accuracy, latency and cost as real usage comes in.