commis — self-optimizing restaurant marketing from a single video
Inspiration
Small restaurant owners are busy cooking, not learning how to run a brand agency. What if you could film a quick walkthrough and automate everything from there?
What It Does
Upload a walkthrough video. Commis extracts the restaurant name, full menu with prices, photos, cuisine type, and vibe. It generates and deploys a live website, then improves it autonomously via A/B testing.
It also cuts cooking clips into YouTube Shorts with AI voiceovers and subtitles, generates social graphics, recommends Google Ads campaigns, and aggregates reviews into digests. A voice assistant controls everything. Owners give minimal input and get good output.
How I Built It
Gemini's native video understanding was the biggest win over frame extraction. Frame-by-frame caught maybe 5 menu items. Full video with two-pass Gemini Pro gets dramatically more, understanding spatial context and reading background signs.
Website generation uses an iterative loop: Gemini Pro generates HTML, Puppeteer screenshots it, a UI evaluator which uses Gemini along with deterministic tests scores it, and feedback drives another pass if needed.
A/B testing uses Thompson Sampling, a Bayesian approach that converges faster than frequentist methods at low traffic. Winners auto-graduate at 95% confidence to be served as default through the main URL and new experiments queue autonomously through a cron every 4 hours.
Everything connects through an MCP server with 14 tools, working across voice assistant, web UI, or AI agent. Backend is Express + SQLite, video processing with FFmpeg/Sharp, deployment to Cloudflare Pages.
Challenges
Iterative generation required solving a feedback problem: Gemini needs to know why a site scored poorly to fix things. I built a remediation system mapping common failures (broken mobile nav, bad contrast, tiny touch targets) to specific prompts for retries. For more rare ones, the Gemini-powered evaluator itself comes up with a prompt to fix.
What I Learned
Iterative AI workflows dramatically beat one-shot generation. Letting the model critique and retry costs more but produces far better output. Thompson Sampling is great for small-scale optimization.
Log in or sign up for Devpost to join the conversation.