Tipfork

Ordering at a restaurant is hard when the menu is in another language, dishes are just names with no pictures, and the end-of-meal math (tax, tip, who owes what) lands on one stressed person. TipFork solves the whole flow in one mobile-first app. Point your camera at a menu and TipFork's agent pipeline takes over. A Fireworks AI vision-language model (Kimi K2.6) reads the photo directly and extracts structured dish names and prices, with local Tesseract.js OCR as a merge-and-fallback signal. One tap translates every dish into your language using gpt-oss-120b on Fireworks, tuned to keep names concise and food-accurate. Another tap generates a visual of each dish (FLUX.1 schnell via Fireworks, with a Qwen image fallback), so you know what you're ordering before it arrives. Then TipFork handles the money math without ever touching money: it estimates sales tax from a built-in regional rate table (auto-detected via geolocation or picked manually), applies your tip choice, and splits the bill evenly or dish-by-dish per person. After the meal, photograph the receipt and TipFork reconciles the actual charged total against its estimate, adjusting each person's share so what you collect matches reality. Engineering highlights: every AI route degrades gracefully (OCR fallback for extraction, labeled passthrough for translation, styled placeholders for visuals), image generation runs with bounded concurrency, per-item timeouts, an LRU cache and a route-level time budget, and the whole app installs as a PWA with Capacitor packaging for iOS/Android. All AI inference runs on Fireworks AI serverless — approved compute for this hackathon — behind a single documented Node/Express backend.