Skip to content

AIProductSep 2025 – Feb 2026

Fixr — Vehicle Maintenance Tracker

A vehicle maintenance tracker with a Gemini-backed assistant that explains services in plain language, and the verification layer it still lacks.

Next.js · Vercel · JavaScript · Convex · Gemini API · NextAuth · Resend · Tailwind CSS

Fixr dashboard with vehicle health score and computed maintenance suggestions
What it is
A vehicle maintenance tracker with Fixy, a Gemini-backed assistant that explains service needs in plain language and estimates what they should cost.
Hardest problem
Users spend real money on the answers, and nothing currently sits between the model’s text and the screen — no source consulted, no claim scored. The maintenance schedule table that could check the numbers already exists in the repo; it just feeds a separate non-AI path the chat action never imports.
What I'd change
Wire the existing schedule table into the chat path: have Fixy return figures as structured data, compare them in code, and mark each claim in-range, contradicted, or unverified next to the answer. Then OCR for receipts and OBD-II for mileage.

The Problem

Vehicle maintenance is confusing and expensive. Most people don't know when to service their car, what maintenance actually means, or whether they're being overcharged by mechanics. Missed maintenance leads to costly repairs; unnecessary maintenance wastes money.

Existing solutions are either too technical (maintenance manuals), too generic (calendar reminders), or vendor-locked (dealership apps). People need a personal maintenance assistant that speaks plain English and keeps their specific vehicle history.

My Approach

I built Fixr as a personal vehicle maintenance companion that tracks service history, sends reminders, and answers questions through "Fixy," a chat assistant that explains maintenance in everyday language.

The app uses a dashboard-first interface showing a vehicle health score, pending tasks, and maintenance suggestions computed from a schedule table against mileage and service history. No model is involved in that path. Each suggestion carries a cost range and the interval it came from.

A mechanic finder built on Google Maps shows nearby shops with ratings and reviews. Fixy explains technical terms in plain language. As the summary above says, nothing checks its answers.

What it looks like

Fixy answers with the vehicle’s make, model and mileage in context. The figures in an answer like this are the model’s, unchecked against the maintenance schedule table that already exists in the same repo — the gap described above.
Tasks are automatically prioritized by urgency. Upcoming tasks include mileage progress bars showing how close the vehicle is to the service threshold.
Users can create custom maintenance reminders with both date and mileage triggers. Priority is auto-calculated but can be manually overridden.
Logging a completed service captures provider, cost, mileage, and an optional receipt upload to build a full verifiable maintenance history.

Technical Deep Dive

Why Convex over Supabase or Firebase? Convex provides real-time reactivity with a better TypeScript experience. Queries automatically update when data changes, perfect for maintenance reminders that depend on mileage and time. The reactive queries mean no manual cache invalidation.

Why Gemini API? Honestly, it was the model I had free API access to, and a sentence of plain-language explanation is not a hard generation task. I don’t have a benchmark putting it ahead of the alternatives here, and claiming one would be inventing it. The system prompt supplies vehicle context and asks for cost awareness, a safety-first framing, and hedging when unsure — which shapes the output but enforces nothing, because nothing checks the result.

The reminder system uses Convex scheduled functions that run daily, checking each vehicle's maintenance schedule against current mileage and time since last service. Resend handles email delivery with customizable notification preferences.

NextAuth with database sessions provides secure authentication. I chose the Credentials provider for email/password auth, storing hashed passwords in Convex. Session tokens are HTTP-only cookies to prevent XSS attacks.

The Google Maps integration uses the Places API to search for mechanics within a configurable radius, filtering for high ratings and returning detailed information including hours, phone numbers, and user reviews.

Challenges & Lessons

Challenge: AI hallucinations about maintenance costs and schedules — still open, and the one I’d fix first. What exists is prompt instruction telling the model to hedge on specs it isn’t sure about, which is the model policing itself. An unenforced instruction is a preference, not a constraint, and I shouldn’t have been counting it as one. I also made it worse once: an early prompt told the model to defer to the owner’s manual when unsure, and I replaced that with an instruction to give specific figures when confident — solving a vagueness complaint by removing the caution instead of adding a way to check. The disclaimer that does exist only renders on the empty state, so it disappears the moment an answer appears. Lesson: a safeguard you can’t point at in code isn’t a safeguard.

Challenge: Users input mileage inconsistently (some regularly, some sporadically). Solution: Added mileage estimation based on time and average driving patterns, with clear UI indicators when using estimated vs. actual mileage.

Challenge: Email deliverability for reminders. Solution: Switched from SendGrid to Resend, implemented proper SPF/DKIM records, and added a digest option to reduce email volume (weekly summary instead of individual reminders).

If I rebuilt this, I'd add OCR for service receipts (snap a photo to auto-populate service details) and integration with OBD-II readers for automatic mileage tracking.

More of the same kind of work

The reasoning behind this project’s central decision is in I Keep Choosing the Less Impressive Model.

Eight projects, four with a model in the loop.

All projects