AI Product Design · Seed → Series B
Your AI feature works. Getting users to trust it is the hard part.
AI features fail differently than normal software — silently, confidently, and in ways users can't always spot. I design the trust calibration, failure states, and human-in-the-loop controls that make AI features something people actually rely on.
Problems I solve.
What usually goes wrong with AI features
- Trust — no signal for confidence or sourcing, so users either over-trust the output or dismiss the feature entirely
- Latency — no feedback while the model "thinks," so the product feels broken or frozen
- Failure — hallucinations and wrong answers have nowhere to go; one bad output erodes trust in every output after it
- Control — nothing lets people edit, override, or correct the AI's work
- Adoption — without a clear path back to human control, people quietly stop using the feature
How I approach it.
Map the failure modes
Catalog what happens when the model is wrong, slow, uncertain, or silent — before designing the happy path.
Design trust calibration
Confidence indicators, sourcing, and explainability, so people know when to rely on the output and when to check it.
Build the human-in-the-loop
Edit, correct, override — every AI feature needs a clear path back to human control.
Validate with real usage
Ship narrow, watch edit and correction behavior, and use that signal to improve the model and the interface together.
The pattern across every AI feature I've shipped.
Four different products, four different teams, one repeated decision: don't let the AI act past what it's actually grounded in — and make that boundary visible to the person using it.
Where that shows up
- BeeTheData — face-match alerts lead with the confidence percentage, not just the match, because that's what the officers acting on them said they needed.
- Otters.ai — the JVP generator is constrained to the data explicitly shared in the source documents. It doesn't invent anything.
- King — an internal GPT for looking up level design data was tightened to answer only from the source spreadsheet, after an early version returned numbers that didn't exist.
- Oraion — the platform's data-building agent auto-approves only above a 0.9 confidence score; below that, it routes to a human, down to "a human writes it from scratch."
Outcomes, not opinions.
Founding Designer
AI-generated Joint Value Propositions rated equal to or better than manual ones by 8 of 9 users, with 88% feature adoption.
Design Lead
Legacy internal tool with significant UX debt and no AI readiness. Led the redesign used daily by designers, developers, and operators across titles like Candy Crush.
Head of Design
Designed the platform's five-layer architecture end to end, including the Lineage view — a full data-provenance graph showing every rule applied and whether a human or the AI approved each step, with timestamps.
Design & Team Lead
AI facial-recognition alerts for the Spanish police. The demo I built is what got it funded, not a pitch deck. After talking to the officers who'd act on alerts, redesigned the UI to foreground match confidence.
First Designer
Built the design system, marketing site, and admin product from scratch. Designed the AI chat to hold each creator's own tone consistently, with a hand-off point back to the real creator as engagement grows.
Co-Founder, Product & Design
An AI agent that fact-checks claims forwarded from social apps. Every verdict is source-cited, with a plain true / misinformation / unclear label instead of a manufactured confidence score.
If you're looking for an expert who understands your needs and can answer them simply, contact Jorge.Rémy Voet · Head of Design, DIGITALinkers
Questions.
Do you build the AI model, or just design the interface around it?
I design the interaction layer — prompting UX, output presentation, trust and error states. I'm not an ML engineer; I work alongside your ML team or API provider, not instead of them.
What if we haven't decided which AI features to build yet?
I can run a short discovery pass to find where AI actually adds value — and where it doesn't — before any design work starts.
How do you design around hallucinations and wrong answers?
By assuming they'll happen. Every AI feature gets a plan for low-confidence output: sourcing, confidence signals, and an easy path to correct or override the result.
Does this replace our prompt engineer or ML team?
No — it complements them. I focus on how the AI's output reaches the user; prompt and model work stay with your technical team.
What does "human-in-the-loop" actually look like in the UI?
Edit affordances on AI output, confidence indicators, an obvious way to reject or regenerate, and a clear record of what was AI-generated vs. human-approved.
Can you work with an AI feature we've already shipped?
Yes — I audit what's live, find where users are confused or dropping off, and redesign around actual usage and edit patterns.