Specialized engagement
Ship an AI feature people actually use
AI features die two ways: a demo that never ships, or a shipped thing nobody trusts. A focused sprint to design and build one AI feature into your product — the interaction, the prompts, the failure states, the eval — not a chatbot bolted to the corner.
From $30,000
2–3 week fixed sprint. One feature, shipped.
What you get
- The interaction designed end to end — input, streaming, edit, retry, and the empty and error states
- Prompt and tool design with an eval harness, so quality is measured, not vibes
- A trust layer: citations, confidence, undo, and human-in-the-loop where it matters
- Working production code in your stack (Next / React), not a Figma file
- A latency and cost budget, so the feature is shippable — not just impressive on stage
Why this, from me
I design the UX and write the code and the prompts, so the model's behavior and the interface are one decision. Most AI features fail at the seams between design, prompt, and engineering — there are no seams when one senior owns all three. The tools on this site are AI-shaped; I build with the latest Claude models every day.
- Design, prompt, and production code owned by one senior
- Evals before ship — quality has a number, not an opinion
- Failure states designed first, because that's where AI features lose trust
How it runs
01
Fit call
30 minutes to pick the one feature worth shipping and the failure modes that matter.
02
Map
The interaction, the prompt strategy, and what 'good output' means — written down as evals.
03
Build
UX, prompts, and evals built together against a latency and cost budget.
04
Ship
Production code in your stack, with the eval suite so you can keep it honest.
Who it's for
A fit
- A funded product adding its first real AI feature
- A team with a demo that won't survive real users
- Founders who want it shipped, not slideware
Not a fit
- Pure research with no product to ship into
- A full AI platform — that's a team, not a sprint
Start with a fit call
Tell me what you have in mind. If it's a fit, we book a call and I scope it. If it isn't, I'll point you somewhere better.
Questions
The latest Claude or GPT models, chosen per task. I build with them daily and pick by what the feature actually needs.
A test suite for model output quality — run before ship so 'is it good' has an answer, not a shrug.
One feature, clear scope. The fit call sets the number before anything starts.