Le Pathétique
BearHacks 2026 • AI Food Critic
Notice: A Vercel deployment exists but is still being cleaned up after the hackathon, so the DevPost entry is the source of truth.
About this project
The problem
Hackathon judging happens in a loud room on unreliable Wi-Fi, and most AI demos die the moment an API call hangs. I wanted something genuinely fun that would still work if the network or an API key failed mid-presentation.
My role
Solo build over one weekend. Persona design, the Vision to Gemini to ElevenLabs pipeline, the interface, and the offline fallback mode.
What I shipped
- Chained three separate AI services into one pipeline that turns a photo into spoken audio critique and three tiered recipes.
- Built an offline mock mode that keeps the full demo working when the venue Wi-Fi or an API key fails, which is exactly what a live demo needs.
- Wrote 6 distinct critic personas with 13 hidden easter eggs so repeat use keeps finding new material.
- Went from empty repo to working multimodal demo in a single weekend, solo.
Achievements
4 of 4 unlocked
Solo Hack Shipped
Built, shipped, and demoed Le Pathétique at BearHacks 2026, solo, in one weekend.
Multimodal in One Shot
Gemini Flash handled vision and creative writing in a single call. No chaining required.
Wi-Fi Will Not Stop Me
An offline mock mode keeps the demo alive when venue Wi-Fi or an API key dies.
13 Easter Eggs
Six critic personas, thirteen hidden easter eggs. Repeat-play tested.
Case study
Le Pathétique (BearHacks 2026)
DevPost: devpost.com/software/le-pathetique GitHub: github.com/Cengizbey-m/Le-Pathetique
A pretentious French food critic reviews your cooking out loud, then hands you three recipes to redeem yourself. Built solo at BearHacks 2026 over one weekend.
The pitch
Upload a photo of whatever sad meal you are looking at. The app works out what is actually in it, then a French-accented voice delivers a critique somewhere between disappointment and fury. Then it offers a way out: three recipes tagged broke, decent, and fancy, so you can fix dinner at whatever budget you have.
The pipeline
- Photo upload in the browser.
- Google Cloud Vision labels what is in the picture, so the model does not confidently review a bouillabaisse when you are holding a granola bar.
- Gemini Flash takes the labels and the image and writes both the critique and the three recipes in a single multimodal pass.
- ElevenLabs synthesizes the critique in a French accent.
- Web Audio API layers in sound effects generated in the browser, with no audio assets to download.
The decision that saved the demo
Hackathon Wi-Fi is a coin flip and API keys have a habit of failing at 2 AM. I built an offline mock mode: if Vision or Gemini fails, the app falls back to canned but plausible critiques and recipes and the demo continues. Judges never see a spinner that does not resolve.
That was the right call for the same reason it would be in production. A user-facing feature that depends on three third-party services needs a defined behaviour for when one of them is down.
Other things I am happy with
Six personas, 13 easter eggs. Repeat visits surface different critics with different voices and hidden interactions, which gives the demo replay value instead of one laugh.
A deliberately ugly interface. Most hackathon projects share the same look: Tailwind defaults, soft shadows, gradient buttons. I went the other way with serif headlines, hard borders, and monochrome with a single ink-red accent. It stood out on a table of forty similar-looking laptops.
One call instead of two. Gemini handles the visual analysis and the creative writing together, rather than chaining a separate vision model into an LLM. Fewer moving parts, less latency, fewer failure modes.
Tech
Next.js 16 (App Router), React 19, TypeScript, Tailwind v4, Google Cloud Vision, Gemini Flash, ElevenLabs, Web Audio API, Vercel.
Status
Built solo at BearHacks 2026. A Vercel deployment exists but is still being cleaned up after the event, so the DevPost entry is the source of truth. The source is public.