Captured from the live deployment, not a mockup
01 · The problem
Ranked #12 of a 24-hour agentic challenge by letting the agent propose and a deterministic layer decide: a 90-day cash-flow forecast keeps only payment plans that never breach the minimum balance.
02 · How it works
- 01
Re-applied the deterministic ranking after every model call so an LLM choice cannot silently alter a graded field, and verified each rendered row against the submission contract with self-repair, emitting the next safe plan when the top one fails.
- 02
Made the whole run reproducible from cold: unzipping the submission with every AWS variable cleared regenerates the output byte for byte, so the result does not depend on the machine that produced it.
- 03
Built a financial decision agent with typed evidence extraction on Amazon Bedrock and a deterministic 90-day cash-flow forecast, ranked #12 in a 24-hour agentic AI challenge.
03 · What it cost, and what it returned
Verified every rendered row against the submission contract with self-repair, and made the run reproduce byte for byte from cold with every AWS variable cleared.