Dual-mode AI automotive platform. GPT-4o + Claude in parallel. 4 weeks.
Canadian car buyers were making $30,000–$60,000 decisions with almost no reliable intelligence. Dealers were spending 45 minutes per vehicle on manual pricing research. Existing tools were either too shallow (consumer-grade) or too expensive ($500–$2,000/month dealer systems). Nothing served the gap.
Auto Detective has two distinct product modes served from one Base44 app. Consumer mode is publicly accessible — any visitor can generate a vehicle report by entering a VIN or pasting a listing URL. Dealer mode requires a subscription and is gated behind authentication and a plan check.
The AI analysis pipeline is the core architectural component. A report request goes to a backend function that: (1) looks up VIN data, (2) fetches comparable Canadian listings, (3) passes the assembled context to GPT-4o for the market pricing report, (4) passes vehicle history and notable flags to Claude for red flag analysis, (5) assembles both outputs into a structured Report entity. Running GPT-4o and Claude in parallel (Promise.all in the backend function) rather than sequentially reduced report generation time from ~8 seconds to ~4 seconds.
The VehicleReport entity stores the full AI output — raw GPT-4o pricing analysis, Claude red flag summary, and final assembled report. We store the full JSON so we can regenerate the formatted report without re-running AI. Report generation cost (API fees) is tracked in ReportUsage per dealer per month so we can enforce plan limits and show dealers their usage.
Auto Detective uses GPT-4o and Claude in parallel, each doing what it does best. This was a deliberate architecture decision made after testing both models on the same tasks. GPT-4o produces cleaner market pricing summaries and writes better plain-English recommendations. Claude handles long vehicle history documents and red flag reasoning better — it's more reliable at nuanced analysis when given a long context with ambiguous signals.
The backend function that runs the analysis: (1) fetches VIN decode data and comparable listings, (2) assembles the full context string (VIN data + comparables + vehicle history), (3) fires Promise.all([invokeGPT4o(pricingPrompt), invokeClaude(redFlagPrompt)]), (4) waits for both to resolve, (5) merges the outputs into the report structure. API costs per report are low at typical usage volumes — well within the margin of the subscription tiers.
The AI agents for buyer Q&A are configured on Base44 with read access to the VehicleReport and DealerInventory entities, plus a backend function tool that generates fresh comparables on demand.
Dealer subscriptions use Stripe's Subscription API with two plans: Professional ($79/month) and Enterprise ($199/month). Trial management was a key implementation detail — dealers get 14 days free, and the trial end date is stored on the DealerProfile entity. A daily scheduled automation checks for trial expirations and sends reminder emails at day 10 and day 14.
The subscription checkout uses Stripe Checkout with a subscription mode session. On checkout.session.completed, we create the DealerProfile and Subscription entities and send a welcome email with login instructions. Subscription cancellations trigger a scheduled deletion — the account stays active until period end, then a webhook event sets the dealer's plan to 'cancelled' and restricts report generation.
Canadian automotive listing data from different sources uses inconsistent formats — some prices include taxes, some don't; mileage is in km or miles depending on source. We built a normalization layer in the backend function that standardizes all comparables before passing them to GPT-4o. Without this, the AI pricing analysis was unreliable.
Running GPT-4o and Claude in parallel via Promise.all works well, but if one fails, the whole report fails. We wrapped each AI call in an independent try/catch — if Claude's red flag analysis fails, the report still generates with GPT-4o's pricing output and a 'red flag analysis unavailable' note.
Before launch, we ran a batch of historical vehicles through the AI pipeline and compared AI pricing to actual sale outcomes. The large majority landed within an acceptable range. Outliers were unusual vehicles — modified trucks, rare trim levels — where comparable data was sparse. We added a confidence score to the report (high/medium/low) based on comparable count.
The full case study covers results, metrics, and what the client said.
Free 30-minute call with Ryan. Fixed quote before we hang up.
$100/hr CAD all-inclusive. 2–6 week delivery on Base44.
Book Free Discovery Call