Five repos I keep returning to: NCNN on Pi 3, Roboflow Universe, Turso libSQL, Pagefind

Description and use

Low‑ceremony solutions for AI inference on Raspberry Pi, search, and data storage that eliminate GPU, per‑query charges, and cloud dependencies.

Problem

Deploying AI and search solutions on constrained hardware and free-tier services without incurring GPU costs, per-query charges, or complex cloud dependencies.

Outcome

Achieved low‑cost, GPU‑free inference on Raspberry Pi, obtained commercial‑licence datasets without manual labelling, leveraged free‑tier edge SQL, and implemented static search without runtime API calls, all within constrained hardware and budget.

Workflow

• 1. Export YOLO11n model to NCNN format and run inference on Pi 3, adjusting input resolution to fit memory.
• 2. Train YOLO11n on Apple MPS, evaluate on held‑out set, and export to NCNN.
• 3. Search Roboflow Universe for CC BY 4.0 datasets, download, audit licences, and merge into a single dataset.
• 4. Design Turso schema within free‑tier limits, deploy on Turso, and use TypeScript SDK for queries.
• 5. Integrate Pagefind with Astro, generate WASM index at build time, enable lazy‑loading and native dialog UI.
• 6. Deploy sites on Vercel, ensuring static rebuilds keep search index fresh.

Evidence

NCNN doesn't assume a GPU and has no external runtime dependencies.

Source

I merged 7 Roboflow Universe datasets — 11,667 images total, selecting only datasets with CC BY 4.0 licences.

Source

Turso free tier has row limits and database count limits shaped the data model more than the application logic did.

Source

Pagefind generates a search index at build time — no external service, no API key, no monthly bill after the first day.

Source