A second devpost was accidentally made and information is split across both. They were both submitted by the deadline.
What it does
Underight runs a commercial loan from the first document to funding. Reads every file as it lands. Applications, IDs, formation docs, bank statements, and tax returns are structured as fields and checked for consistency. Moves the deal through each stage. Intake, Credit, Committee, Product, Committed, Closed. Missing items become borrower requests. SLA timers send alerts naming the deal, the stage, what it's waiting on, and what unblocks it. Keeps people in charge. Credit, committee, sign-off, and funding decisions go to the right person with a short summary. Underight requests fraud holds; a human places them. Compares every deal with every past deal. A new $248K equipment loan passes all five existing checks. Underight sees its vendor remit account, bookkeeper, and registered agent address match two funded deals, and that the real vendor pays to a different account. It requests a hold before money goes out and links the earlier deals. Logs every decision, so any hold can be explained.
How we built it
A Python event-driven engine runs on a simulated clock that replays how files arrive during a real week. An LLM extracts fields from each document. Versioned rules handle stage requirements and the five known fraud checks. A fingerprint index over vendors, remit accounts, preparers, agents, addresses, ID numbers and PDF metadata finds links across deals. Everything goes to an append-only log and is exported as scored JSON.
Challenges we ran into
Speed and safety pulled against each other. A fast lane for clean-looking deals would have sped up the fraud. We grant trust based on verified facts (remit account matches the known vendor, prior relationship history), not on how tidy the file looks.
Accomplishments that we're proud of
Caught a ring that passed every existing check, using signals already in the bank's own data. Every alert names an owner, a wait time, and a next step. No decision happens without a person. What we learned Fraud rings reuse the same setup even when every document is clean: the same bookkeeper, agent, payout account, and PDF tool. One deal on its own looks fine. Compared against the rest of the bank's deals, the pattern shows.
What's next for Underight
Live bank account checks through Plaid, Secretary of State and ID record checks, alerts sent through the bank's own messaging tools, and an AI red team that keeps designing loan files to beat Underight, so each attack that works becomes a new check.
Github: https://github.com/mnettem21/DealPilot