Show & Tell: AI-powered recipe generator from fridge photos
I built FridgeChef — take a photo of your fridge contents and get recipe suggestions! Uses GPT-4o vision to identify ingredients, then generates recipes.
Tech stack:
The vision model is surprisingly good at identifying specific ingredients, even partially obscured ones. It correctly identified 87% of items in my testing.
Live demo at [link]. Built this over a weekend hackathon. Would love feedback!
For medical applications specifically, I'd add these strategies:
5. Citation verification: Require the model to cite specific paragraphs from source docs 6. Factual extraction only: Don't let the model generate — only extract and restructure 7. Multi-model consensus: Run the same query through 2 models and flag disagreements
We use strategy #7 in our research pipeline and it catches ~60% of hallucinations that single-model validation misses.
Multi-model consensus is clever but doubles the cost. We implemented citation verification and it's been our most effective single strategy. If the model can't point to a specific passage in the source document, we flag it for human review.
This is an area we're actively working on. Some tips from our end:
- Use lower temperature (0.1-0.3) for factual tasks
- Structured Outputs can constrain the output to valid options
- The Assistants API with file_search automatically includes citations
We're also working on built-in hallucination detection features. Stay tuned!
Log in to reply to this topic.