Photograph a leaf and see the likely disease plus treatment steps
Intermediate20h2–4
PythonYOLONext.js
16/20Few-Shot Image Dataset Labeler
Label 20 images and let the model suggest labels for the next thousand.
Advanced28h2–3 people16/20 fit
Build Few-Shot Image Dataset Labeler for ml students and kaggle competitors in ~28h with 2–3 people. Stack: Python + CLIP + Embeddings + Next.js.
Why it wins: Active learning is a real ML idea, demoed in a visual way.
Build flow
01 · Problem
Labelling image datasets by hand is the slowest part of every computer-vision project.
02 · Solution
Embed images with CLIP, train a light classifier on a few labels and use active learning to surface the most uncertain images for the human next.
03 · Build
- 1Embed images with CLIP
- 2Train logistic regression on labels
- 3Rank unlabelled images by uncertainty
- 4Export labels
04 · Demo
Label 10 images live and watch the suggestions get better each round.
MVP checklist
- Few-shot labelling
- Uncertainty queue
- Accuracy per round
- Export to CSV or COCO
Tech stack
| Tech | Role | |
|---|---|---|
| Python | Backend / scripting | Backend |
| CLIP | Image-text embeddings | AI |
| Embeddings | Semantic similarity | AI |
| Next.js | Frontend + API routes | Web |
Related ideas
More AISearch your photos by description without uploading them anywhere
Advanced30h1–3Transformers.jsCLIPWebGPU17/20Incoming tickets get a category, urgency and suggested reply automatically
Intermediate14h1–3EmbeddingsLLMPostgres16/20