# Sashi Perera

- **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:** [Sashi Perera](https://cerebralvalley.ai/u/sashiperera)
- **GitHub:** https://github.com/WAOmaster/bidledger
- **Demo video:** https://github.com/WAOmaster/bidledger/releases/download/v1.0/bidledger-demo.mp4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5325

BidLedger: every shouted bid, on the record.

The problem: across India and Sri Lanka, farmers' fresh produce is sold by open-cry auction at wholesale markets. An auctioneer calls a lot and traders shout bids, often in different languages. The only record is the commission agent's handwritten slip, and the farmer is paid from that slip. If the slip shows a lower price than what was shouted on the floor, the farmer has no way to prove it.

What BidLedger does: one phone records the auction floor. BidLedger turns that recording into a verified sales ledger with audio proof:
1. Muse Voice Transcribe in DIARIZATION mode separates the auctioneer and the traders and timestamps every turn (English, Hindi and Tamil language bias plus auction keywords). An optional denoise step cleans the floor recording first.
2. Muse Spark 1.3 acts as the auction clerk: it works out which voice is the auctioneer, which trader is behind each voice, reads bids spoken as Hindi or Tamil numbers, and returns every lot's bids, winner, price per kg and the turn where the sale was declared.
3. BidLedger cuts the audio proof clip of each winning call.
4. It checks the commission agent's slip against the floor and flags any shortfall to the farmer in rupees.
5. Each sale becomes two WhatsApp Cloud API utility messages, one to the buyer and one to the farmer, with the proof clip.

Results on the live Meta Model API: 4/4 lots with the correct buyer and price, and a Rs 9,000 underpayment on lot 1 caught with the clip to prove it. We stress-tested it under REAL market recordings from Lagos, Gombe, Cebu and Huancayo (Wikimedia Commons): 4/4 at 5 dB and at 0 dB, where the market is as loud as the bidders; and on synthetic crowds down to -1.5 dB, 4/4 every time. On a real auctioneer's chant with no lot or winner called, Spark returned no sale instead of inventing one (NOISE-TEST.md).

How Meta earns (submission/META-BUSINESS.md): daily Model API usage per market; 2 paid WhatsApp utility receipts per lot (India alone: 1,522 e-NAM mandis, roughly 1.5M messages a day at scale); a pay-the-farmer button on WhatsApp Pay (UPI, Pix); Click-to-WhatsApp ads from agents and buyers; a Business Agent that answers "what did my carrots sell for?"; Ray-Ban Meta glasses for hands-free recording. The same engine fits fish landing auctions, cattle fairs and day-labour sites.

Honest note: bidder voices are synthetic so every lot has a known true result; video images were generated with Muse Image. Everything was built during the hackathon. Next: a live market floor, realtime receipts as the hammer falls, and SAM 3.1 to track raised hands for silent bids.

## More from Meta Global AI Developer Hackathon

- [Whatauth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5322)
- [Videonerd](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5323)
- [Proof Loop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5324)
- [Kumud Waykole](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5326)
- [zorith](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5327)
- [prompt2brand](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5328)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5325. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt