Daniel Kitchen.
Twenty years turning a customer's problem into a shipped, working system. Now I build those systems with AI agents — in production, measured, and trusted to run on their own.
For twenty years I've done the forward-deployed job without the title: sit with the problem, scope it, build it, and own it until it works. At The Diffusion Co. I drove products from a first discovery call through twenty-three prototypes to paying customers — coordinating outside engineering contractors I had no authority over. The technology kept changing; the shape of the work never did. Today it's LLMs and agent systems, shipped with real evals.
AJO.
A message bus that runs a fleet of Claude agents as one system — 18 projects, 85 roles, one router. Discovery, the Pain Point Pipeline, Researcher, Tool Engineering and more all live and coordinate inside it. Swap the objective; the coordination underneath stays the same.
See the router in action →Discovery
The never-settled search — reformulates how the fleet finds what's worth solving, and stands up focused engines to do it.
Walkthrough →Pain Point Pipeline
Discovery's live engine — one query to a cited pursue / hold / pass on a buildable pain, grounded in real spend.
Walkthrough →Researcher
Ranked, quality-gated research questions scored by value of information. Pairs with the Pipeline into a self-driving loop.
Walkthrough →Tool Engineering
An autonomous factory that carries a qualified pain to a shipped, tested tool — a human only at four gates.
Walkthrough →Agent Dev
Builds the roles the fleet runs on — a contract plus a verification loop, not a persona. It forges fixes from real failures.
Walkthrough →Systems
Keeps the fleet running unattended — monitoring, incident response, and the guards that make autonomy safe.
Walkthrough →Tracking
The seam between the fleet and me — a live queue of every job, plus the Netlify dashboard that puts all of AJO in my pocket.
Walkthrough →Shipped · live in the world
4 productsYonder
One new dish at a controlled distance from your routine, with the reason it's worth cooking. A calibrated-novelty engine grounded in 2,476 recipes, on Cloudflare's edge.
Case study →Fieldstone
A support agent that knows when not to answer — five typed tools, dual-signal confidence, a clean escalation path. 98.1% on an Opus-judged golden set.
Case study →Curie
A three-bottle piezoelectric diffuser, concept to 23 paying customers — industrial design, custom PCB, Arduino firmware, native iOS app, retail packaging.
Case study →Lever
Two breathing protocols from published research in one offline-capable page. No accounts, no streaks, no upsell — built in a single session.
Case study →Let's build something that ships.
I'm most useful where an AI agent has to leave the demo and survive contact with real users — deployed, measured, and trustworthy enough to run on its own. If that's the problem in front of you, let's talk.