SIRBHAIYA ISHAAN

Walnut on Switch

Flip the Switch Hackathon/Nov 2026

Regulated teams (finance, fintech, health) can't put a normal AI bot in a shared channel: everyone in the room reads its answer, so one reply can leak restricted material, and nobody can prove afterwards who saw what or who approved an action. If the money-laundering reporting officer's answer mentions a case, the account manager in the same channel has just been tipped off. Walnut on Switch brings a permission-aware company brain into team chat. Two people ask the same question in a shared room; each gets a cited answer built only from sources they may open, delivered to their own private room, while the shared room sees one identical line. For someone without access, a restricted case looks exactly like a company that doesn't exist. Reply drafts and tickets become approval cards: the requester can never approve, and the approver must be able to open every cited source, re-checked live. Every step is sealed in a hash-chained audit ledger that compliance can verify from chat. It's for compliance-heavy teams that already work in chat. Status: tested against Switch's frame format and a local test bridge; not yet run on a live Switch server, and the video uses the test bridge (shown on screen). All data is synthetic (a made-up fintech, "Company A"). Reused pre-existing code: the Walnut Internal Brain and platform; new during the hackathon: the Switch agent client, the chat workflow, the room template and 73 tests (file by file in docs/SWITCH_CHANGES.md). Team: Ishaan Sirbhaiya, Dev Sirbhaiya, Mehul Modi, Aaruni Garg (Walnut, walnut.sg).