# All Clear

- **Event:** [Agents in Context Toronto: a MongoDB Hackathon](https://cerebralvalley.ai/e/mongodb-toronto-hackathon)
- **When:** Sat, Nov 14 at 9:00 AM – 9:00 PM (EST)
- **Where:** Toronto, Canada
- **Team:** [Subodh Thallada](https://cerebralvalley.ai/u/Subodh), [Caesar Zhou](https://cerebralvalley.ai/u/CzAA), [Ryan Qin](https://cerebralvalley.ai/u/Ryan-LC)
- **GitHub:** https://github.com/subodh-thallada/all-clear
- **Demo video:** https://youtu.be/cbkEuFtwZPs?si=hamY5ooRT12LFx1V
- **Gallery:** https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/51

Phone call demo: https://youtube.com/shorts/lbGS0rggzK8?si=p0gZP5freCBqkANK

All Clear is a recall agent with phone call feature that will not say "all clear" until the database proves every recalled unit is off the shelf.

The problem: when a food batch is recalled, the notice reaches stores, but proof that the product is gone rarely comes back. In the 2025 ByHeart formula recall, the FDA found recalled product at over 175 locations in 36 states, weeks later. Today's tools record a tick from each store. Nobody checks the count.

How it works: complaints from the company's form are classified by defect (reason code first, otherwise Vector Search over past confirmed complaints). Three complaints of one defect on one batch from three stores raise a signal, and a person approves the recall. No model can open one. The agent then phones the plant that made the batch, places holds in the distributors' order systems, calls stores, and hands a sub-recall to a chain that runs its own copy of the agent. Every tin must land in one of six buckets (sold, on shelf, held, returned, forwarded, unresolved), and Atlas refuses to close the recall while any tin is unlocated, disputed or still for sale.

Scenario: an invented 800 g milk powder tin with a faulty lid seal. Batch B-4471, 2,400 tins, about 40 parties, 10 simulated days. The run starts with 5 names and finds the other 35 as it goes.

Lie to it, kill it, break the system under it:
- Lie: a store that reports more tins than it was shipped is rejected by an Atlas validator. The voice agent (ElevenLabs over Twilio) reads back the two numbers Atlas returned and asks again. The model never sees the shipped count, so every challenge comes from the database.
- Kill: four workers claim parties under leases with fencing tokens, and a supervisor kills them at random. A worker that adopts a party checks the webhook log and Twilio before it redials, so a phone never rings twice.
- Break: on day 3 a distributor's order system moves to v2. Only in-flight operations are flagged. The model proposes a new mapping, which must reproduce a count the ledger already knows before it goes live. The second distributor repairs from memory with zero model calls.

Proof beats claims: a connected store's own stock record outranks anything said on the phone, a hold counts only when a read-back shows it, and a store's chain keeps its buyer list in its own database while its counts still have to add up in ours.

A different product is a different recall document on the same engine.

## More from Agents in Context Toronto: a MongoDB Hackathon

- [Exactly Once](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/48)
- [GOTIA reversal Agent](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/49)
- [All Green](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/50)
- [Legacy Siege](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/52)
- [Silicon-Valley](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/53)
- [Never Twice](https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/54)

---

Markdown version of https://cerebralvalley.ai/e/mongodb-toronto-hackathon/hackathon/gallery/51. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt