š¤ Latest Submissions
Assay: Multi-Agent Investment Screening Desk
Assay - Multi-Agent Investment Screening Desk Most agents in this hackathon invent numbers. The smarter ones add an auditor that checks whether the AI was honest about its number. Assay checks whether the number is right. Three agents collaborate through Band in a shared room (a "rondaan", patrol round): - Skaut (scout) surveys the data and returns candidates. - Konduktor (conductor) orchestrates, calls a deterministic Python engine as a tool, and writes the reports. - Pengulas (reviewer) re-runs the engine to cross-audit every figure. No agent computes a ratio. The engine is the single source of truth; every @mention handoff, tool_call, and tool_result lands in Band's audit trail. The collaboration runs through Band, not around it. The wedge: the auditor recomputes from source, so it catches a figure that is wrong even when the AI was perfectly honest. Because it never compares to a stored hash, it catches a tampered number and a wrong-when-written number the same way. A hash audit only catches the edit. Business value: put this on a compliance, risk, or internal-audit desk. A figure that looks honest but is wrong clears review, and a financial restatement costs a company about 9% of its share price on average. Only the engine is domain-specific, so the same control fits underwriting, claims, ESG, or tax. The verdict is a pure function: flip one env var and it is byte-identical across brains, framework-agnostic by construction. Demo: the live link (althea.streamlit.app) is the deterministic Audit Bench, pick a company, tamper a figure, watch the verdict flip PASS to FAIL. The full three-agent Band coordination is in the demo video at 1:00-2:11 (tamper-catch at 2:11-3:10). Two brains validated live through Band: GLM-5.2 and Claude. AI directs, code computes, humans decide. Synthetic data, 24 tests green. A screening desk, not investment advice.
19 Jun 2026
