# BanglaGuard

- **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:** [tahsin moin](https://cerebralvalley.ai/u/tahsinmoin)
- **GitHub:** https://github.com/tahsinmoin-debug/banglaguard
- **Demo video:** https://youtu.be/lI22Kp5UXR8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2134

BanglaGuard is a Bangladesh-aware safety agent for Bangla and Banglish chatbots.

The problem: When a Bangladeshi user asks a chatbot for help after a bKash scam, online blackmail or abuse, Meta's Muse Spark Open is safe and helpful, but 18.6% of its replies contained phone numbers that aren't verified Bangladeshi contacts, including Indian helplines like 1930, 1091 and 181. A standard safety guard doesn't catch this, because a wrong phone number isn't "unsafe content", and it turns helpful answers into bare refusals (only 11% helpful).

What I built during the hackathon:
- A tool-calling agent (classify_intent, lookup_verified_resources, verify_numbers, crisis_route) that routes each message as normal, help-seeking, harmful or crisis.
- The model never writes phone numbers. Contacts come only from a hand-verified list of Bangladeshi resources (bKash 16247, 109, 1098, Police Cyber Support for Women, consumer rights 16121, 999), each with its source. A final filter removes anything else.
- A crisis route that skips the model and gives calm steps and 999.
- Replies in the user's own script (Bangla or Banglish).
- An OpenAI-compatible server (change one base_url), a Gradio demo ("DemoPay", a fictional mobile-banking support chat) and a report card.

Results on 1,703 requests (BanglaSafe, MultiJail, 172 legitimate Bengali requests incl. formal journalism, 20 help-seeking cases):
- Unverified numbers: 18.6% (raw) -> 4.2% (standard guard) -> 0.1% (BanglaGuard)
- Helpful redirect: 98% (raw), 11% (standard guard), 86% (BanglaGuard)
- Unsafe: 2.1% / 1.3% / 1.7% (BanglaGuard matches the guard; intervals overlap)
- Over-refusal of legitimate requests: 0% for all

Limitations: one response model, an LLM judge without human validation yet, and helplines need periodic re-verification.

## More from Meta Global AI Developer Hackathon

- [Realroom](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2131)
- [team cheetah](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2132)
- [Kolmo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2133)
- [Tracklab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2135)
- [Chawanvit Vuthikornudomkit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2136)
- [CatPaw Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2137)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2134. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt