Kanha
AI support bot platform where the model runs in your customer's browser. Point it at your website,it crawls pages, generates QA pairs, fine-tunes a small language model, converts it to WebGPU format, and serves it from Hugging Face Hub. Inference happens on-device. No per-query API calls. Flat monthly pricing.
Tech Stack
About This Project
Architecture
Rust backend (Axum + Tokio + SQLx): ~4,000 LOC handling auth, CRUD, web crawling, training orchestration, billing (Stripe + Razorpay), and four background Redis reliable queues (crawl, QA generation, training, email). Auto-detects JS-heavy pages and falls back to headless Chromium. All workers recover from crashes via DB sweep + queue recovery on startup.
Next.js frontend: Marketing site, docs, dashboard, and auth. Tailwind v4, Radix UI, Supabase auth. The playground dogfoods the published npm SDK for inference.
Training pipeline: QA pair generation (OpenRouter/Groq LLM), fine-tuning, MLC conversion to q4f16_1 WebGPU format, upload to Hugging Face Hub. Runs training on fal serverless GPUs.
npm SDK (kanha-ai): Three integration modes: React component, vanilla JS mount(), and <kanha-bot> Web Component. Zero backend dependency at runtime, resolves model artifacts from HF Hub, loads via web-llm, runs inference in-browser via WebGPU.
Read more at: https://www.yatharth.ai/blog/shipping-a-product-alone