# SIkai Saathi

- **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:** [Niranjan Lamichhane](https://cerebralvalley.ai/u/niranjannlc)
- **GitHub:** https://github.com/NiranjanNlc/sikai_saathi
- **Demo video:** https://youtu.be/vG4V6fS9ULU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3768

Sikāi Sathi · सिकाइ साथी 
Your learning companion. Your language. Your pace.

One-liner: A Nepal-first AI companion that reads a Grade 8 student's handwritten math, names the exact mistake, explains it in Nepali, and proves learning with practice — never rote answers.

1. Problem (education sector)
Rote over understanding. Students memorize final answers (e.g. x = 15) instead of learning the step they missed (dividing both sides by 3).
No 1:1 feedback at scale. Teachers can't hand-mark every copy; parents can't always help with algebra.
Language gap. Most AI tutors explain in English. A Nepali-medium student thinks in Nepali (पक्ष, भाग गर) but is tested in a second language.
Phone-first, low-bandwidth reality. The student device is a low-to-mid Android phone on patchy 3G/4G — heavy apps and full-photo uploads fail.
Untrustworthy AI. Generic LLM tutors hallucinate correctness and invent mastery. In education, a wrong "Mastery: YES" is worse than no feedback. 

2. Solution
Submit → Diagnose → Learn → Practice — one learning loop, 4 screens, 3 taps to practice (down from 7 screens / 5 taps in the audit).

Step	Student does	App does
1. Submit	Photo of copy or typed steps	Pinned equation bar; Photo/Type tabs, "Try an example"
2. Diagnose	Reads flagged steps	Highlights wrong line, names misconception in student language ("Look at line 3…", लाइन 3 हेर)
3. Learn	Reads + answers inline	Nepali/English explanation, quick-check with self-check, balance-scale visual tipped by their own answer
4. Practice	Solves fresh variant	Deterministic new question every time, hint → retry (2 tries) → worked solution, "N of 3" progress + activity timeline
Core rule: correctness, misconception labels, and mastery are decided only by deterministic Python (backend/app/math_engine.py). The LLM only explains. Typed steps are diagnosed deterministically even when an LLM provider is active.

## More from Meta Global AI Developer Hackathon

- [Talking Tactile](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3765)
- [CaseAlly](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3766)
- [Say Again](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3767)
- [FollowAlong](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3769)
- [Ripple handoff](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3770)
- [BOLUS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3771)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3768. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt