Parking Percent

Parking Percent is a SaaS platform that turns ordinary parking lot photos into occupancy analytics. Upload a photo from a security camera or drone, and the system returns vehicle counts, occupancy rates, and type breakdowns within seconds — providing the data that traditionally requires expensive IoT sensor networks or manual counting.

Technical Approach

The detection pipeline uses Facebook's Detectron2 with a Faster R-CNN X101-FPN backbone, chosen for accuracy over speed since processing is asynchronous. The model runs on NVIDIA T4 GPUs via Modal.com's serverless infrastructure — containers spin up on demand, process the image, and shut down, keeping GPU costs under a penny per image. Input resolution is upscaled beyond COCO defaults to improve detection of small, distant vehicles in wide-angle security camera footage.

The full-stack architecture is 100% serverless: Next.js on Vercel, Neon PostgreSQL, Upstash Redis for rate limiting, and S3 for storage. A dual compression strategy handles Vercel's payload limits — client-side pre-compression for upload speed, server-side Sharp processing for security and consistency. Detection coordinates are stored as percentages rather than pixels, enabling responsive bounding box overlays at any screen resolution.

Key Decisions

Serverless GPU inference via Modal.com instead of always-on GPU instances — the product doesn't need real-time detection, so tolerating 5-30 second processing times keeps infrastructure costs at roughly $15/month. Lemon Squeezy as Merchant of Record for global tax compliance without building payment infrastructure. A fail-open design philosophy where Redis rate limiting and storage quota checks allow requests through if the backing service is down, preventing auxiliary failures from cascading into total unavailability.

Screenshots