AI Product Design
The hard part of an AI product is not the chat box. It is the moment a user has to decide whether to believe the machine. I design that moment, and then I build it.
What you get
Interfaces that make uncertainty visible
A model that is unsure should not sound identical to a model that is certain. Most AI products get this wrong, and it is the thing that ends up costing them trust.
Correction paths that are worth the effort
When the model is wrong, the user needs a fast way to fix it, and the fix needs to go somewhere. Corrections that vanish teach users that their effort is worthless.
The states nobody demos
Confidently wrong output. Slow responses. Refusals. The empty state before the model has anything to say. This is where AI products actually live.
How it works
Watch the model fail
I use the product until it does something strange, because that is where the design problem is. The happy path takes an afternoon. The failures take the week.
Design the trust
Evidence, uncertainty, correction, and the path back when the model gets it wrong. Flows first, then screens, in short review loops.
Build it
Production React in your codebase. I write the streaming, the loading, and the error states myself, because those are the design.
Scoped per engagement after a short call. Most start with a fixed price audit of the AI feature that is not landing.
Work behind the offer
Common questions
What is different about designing an AI product?
Normal software is deterministic. The same input gives the same output, and you can map every state. An AI product can be wrong, and it can be wrong confidently, and the user has to decide whether to trust it. Designing for that uncertainty is the whole job.
Do you build the AI, or just design it?
Both, on the front end. I have shipped products with streaming responses, model fallbacks, and dual provider setups. I do not train models, and I will say so rather than pretend otherwise.
We have an AI feature nobody uses. Can you help?
That is the most common version of this work. Users tried it once, something felt off, and they never came back. It is usually diagnosable in about a week, and the cause is almost never the model.
What if our AI product is still an idea?
Then we start smaller than you expect. One flow, one real user job, and a working prototype rather than a deck. I build the thing so we can find out whether it is worth building more of.
Often paired with
Ready when you are.
A 20 minute call, no deck required. I respond within 12 hours.
Book a 20 minute call