# ommspark

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Omkar Mishra](https://cerebralvalley.ai/u/omm12)
- **GitHub:** https://github.com/OmkarMishra56/kisannetra
- **Demo video:** https://youtu.be/WcQ8G_GG1OQ?si=1b0WkvcN6IWfO3J7
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5043

KisanNetra is a camera-first crop-health assistant for Indian farmers who can’t get timely expert advice, especially in their own language. A farmer photographs a sick leaf and gets a clear answer in Hindi, Odia or English, in text and by voice.

How it works: SAM 3.1 outlines the diseased spots and the whole leaf, so the app can say exactly how much of the leaf is affected (for example “about 30%, act now”) and show the highlighted area. Muse Spark looks at the photo and the severity, and writes the diagnosis in simple language. A retrieval layer over official advisories grounds every answer, so advice comes from cited sources instead of guesses. Muse Image generates simple action pictograms (remove infected leaves, wear protection, contact the KVK) for low-literacy users, and each one is reviewed by a person before it can appear.

Built to be safe: pesticide advice can cause real harm, so the app only gives treatment steps when a source supports them. When it is unsure, the photo is blurry, or the damage is severe, it says so and sends the farmer to the nearest Krishi Vigyan Kendra or the Kisan Call Centre (1800-180-1551) instead of guessing. A built-in evaluation script measures how often it gives advice without a source, which must stay at zero.

Built for my world: I’m a student in Odisha, where much of the agriculture advice a farmer can reach is in English or Hindi. Odia support, voice output for low literacy, and a mobile-first design are central to the project, not add-ons.

Tech: Next.js PWA frontend, FastAPI backend, Qdrant vector search, Meta Model API (Muse Spark, SAM 3.1, Muse Image), and Meta’s open MMS-TTS models as a voice fallback when a phone has no Odia or Hindi voice.

## More from Meta Global AI Developer Hackathon

- [SaveBell](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5040)
- [Dragology](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5041)
- [Antonio salvador](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5042)
- [Diash Firdaus](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5044)
- [Second Shift](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5045)
- [Jaya Prakash Merepala](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5046)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5043. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt