Fixit
Point. Diagnose. Fix — and know when to stop.
Point. Diagnose. Fix — and know when to stop.
- 01Camera as the front door, not search
- 02Show confidence — never a bare verdict, never a cold %
- 03When unsure, run a safe test — don’t guess
A 94-screen iOS system on one design system, built around confidence calibration and safety-first deferral
The world bins tens of millions of tonnes of fixable goods a year — not because repair is impossible, but because the knowledge to fix is locked behind a moment of intimidation. Fixit is a camera-first AI repair companion: point your phone at a broken thing and it identifies the item and the likely fault with visible confidence, guides the fix step by step, names the exact parts and tools, and — honestly and visibly — knows when to stop and tell you to call a pro.
E-waste is the fastest-growing waste stream on Earth — 62 million tonnes in 2022, up 82% since 2010, and only about a fifth recycled — and textiles are larger still. When things break, people bin them: around 56% of US homeowners would now rather replace than repair, and roughly one in three have thrown out a broken appliance specifically because they couldn’t find someone they trusted to fix it. The barrier isn’t unwillingness, it’s a single intimidating moment — ‘I don’t know what’s wrong, I don’t know what it’s called, and I’m scared I’ll make it worse.’ Every existing tool makes it harder: iFixit, YouTube and manuals all force a text query you can’t write yet, and the new wave of AI chatbots will confidently hallucinate a fix — which, on electrical, gas or structural work, can injure or kill.
The barrier is intimidation, not laziness
People want to fix things — Repair Cafés prove demand outstrips access. What stops them is not knowing the fault or its name, and the fear of making it worse.
A confident wrong answer is the real danger
Vision models hallucinate, and their ‘confidence’ is token probability, not diagnostic certainty. On electrical or gas work, one confidently-wrong step is a safety incident.
Declining builds trust
A tool that sometimes says ‘stop, get a pro’ or ‘replace this one’ is a tool you believe the rest of the time.
A vision model hands you one confident answer — and it’s often wrong.
Fixit shows a ranked differential with honest confidence, and refuses the fixes that could hurt you.
The core loop — from a broken thing to a fix, with a human always able to stop.
Camera as the front door, not search
The hardest moment is naming the problem, so Fixit removes naming entirely. The centre of the tab bar is a capture button, not a feed — point the phone at the broken thing and the product identifies the item and fault across the everyday long tail (appliances, electronics, clothing, furniture, bikes). Out-breadth the guide libraries on coverage; out-design the chatbots on trust.
Show confidence — never a bare verdict, never a cold %
The result screen always shows a ranked differential (most likely / possible / less likely) with a calibrated-language confidence bar — ‘a hunch’ through ‘fairly sure’ — not a false-precision percentage. Grounded (‘based on the manual for your model’) versus inferred guidance is a first-class, labelled state. This structurally prevents the confidently-wrong single answer that is a vision model’s default failure.
When unsure, run a safe test — don’t guess
If confidence sits in the uncertain band, Fixit doesn’t improvise. It hands the user one cheap, safe, reversible test (‘open the bottom panel and photograph the pump’), then re-analyses and animates the confidence moving. Uncertainty becomes productive instead of scary, and the human becomes the model’s sensor.
Safety as the product, not a disclaimer
When a hazardous class is detected — gas, mains voltage, sealed refrigerant, a microwave capacitor, structural — Fixit routes to a blocking screen with no ‘proceed anyway’ button, only ‘why’ and ‘find a pro’. Declining to help is the feature: false ‘stop’ is cheap, false ‘proceed’ can kill, so caution is deliberately asymmetric.
Honest repair-vs-replace
A ‘worth fixing?’ verdict makes the 50% rule visual — part-and-effort cost against a new unit, weighted by age — and will, when honest, recommend replacing and route to responsible disposal. Advice aligned with the user’s actual interest, including ‘don’t bother’, is what earns trust for the next repair.
- ✕A text query you can’t write yet
- ✕One confident answer — often wrong
- ✕Safety treated as a footnote
- ✕No idea when to stop
- ✓Point the camera — no naming needed
- ✓A ranked differential with honest confidence
- ✓Blocking safety stops, by design
- ✓Refuses danger, routes to a pro
The core loop at a glance — point, analyze, diagnose, plan, fix.
The diagnosis that sells the product — a ranked differential with calibrated confidence, never a bare verdict.
Cook-mode guided repair — one step at a time, the exact parts for this step, a timer when it matters.
The safety hard-stop — for gas, mains, refrigerant or structural work it refuses to coach, and routes to a pro.
3D point-cloud scan — the premium capture path for larger or hard-to-photograph faults.
Parts & tools turned into an orderable, have-it / need-it checklist.
Fixit is a concept, and I’ve framed it as one: no real users, no launch, no shipped metrics. What it shows is a complete, coherent system — 94 screens across 17 flows covering the camera-first core loop, confidence-calibrated diagnosis, the low-confidence guided-test loop, a 3D AR scan, parts and ordering, repair-vs-replace, the blocking safety system, library, account, paywall and the full set of edge and error states — all assembled from one reusable design system, with confidence calibration and safety-first deferral as the signature. The honest test is an expert review: would a repair professional accept that this guides a novice safely, and refuses at the right moments? That’s the bar I designed to.
- Designing for an intimidated novice changed the brief — the most valuable decisions were about restraint, not features: not naming, not guessing, not coaching danger.
- Honesty is a UX pattern. Showing uncertainty, and a safe way to resolve it, built more trust than any confident answer could.
- Regulation-grade safety thinking sharpens a product. Treating the hard-stop as a designed, first-class state — not a footnote — is what separates this from the AI-chatbot wave.
- A 94-screen system forces real systems discipline — tokens and components first, so a confidence reading or a severity dot means the same thing everywhere.
