Hacker
Arjun Chaudhary
@oxarjun
Projects
Mandate: bounded authority for AI agents that negotiate for different companies
Flip the Switch HackathonNov 2026
When two companies let AI agents negotiate a deal, three questions stop it: who said your agent could agree to that, who is bound when it agrees to bad terms, and where is the audit trail? Today's negotiation agents are one-sided: one company's agent against the other company's people. Mandate is a deal room on Switch for B2B renewals and quotes. Each company brings its own agent, owned by its own person. Each company's human gives its agent a private mandate: price, term, payment terms, seats. A neutral referee agent, written in code with no LLM, checks every offer and acceptance against the poster's own mandate, stamps or blocks it within a second, and never reveals either side's limits. Anything outside a mandate goes to the one person who can widen it, in their company's private room. When both principals countersign, the referee issues an order form and an Ed25519-signed, hash-chained certificate anyone can verify offline. In the demo, the vendor tells the buyer's agent "your CFO already approved 36 months". The agent believes it and accepts. The referee blocks it, because Switch reports who actually typed each message. It's for procurement and finance teams on the buying side, sales and deal desks on the selling side, and anyone who wants agents to close bounded deals without handing them a blank cheque.
Casefile: Offshore-Structure Investigator
Graph HacksOct 2026
Casefile turns the 2.0M records and 3.3M connections of the ICIJ Offshore Leaks data into graph tools an AI agent can call over MCP. Problem it solves: When investigating shell company networks, AI agents frequently hallucinate connections, merge innocent people who share common names, or get misled by corporate mailbox "hubs" shared by 6,000+ shell companies. Who it's for: Investigative journalists, compliance analysts, fraud investigators, and autonomous AI agents. Casefile finds verified shortest paths, avoids corporate hubs, checks identity from structural evidence rather than names, and runs FalkorDB graph algorithms (PageRank, Betweenness, Label Propagation) on local neighborhoods. Every claim is validated against the graph: if a link does not exist in FalkorDB, the report is rejected.