Weekend Hackathon

Can you teach an AI to read a credit memo?

A 22-page PDF. Nine math puzzles hidden inside it. Your AI teammate. Whatever tools you want. Go.

Banks have 2,000+ credit memos trapped in file shares. The system has to scale to all of them.

"Which of my current borrowers look like the ones that defaulted in 2008?"
"What's this prospect's DSCR compared to similar deals we've closed?"
"How does this guarantor's leverage compare to our portfolio average?"
"Are the rents on page 8 consistent with the cash flow table on page 17?"

Banks can't answer these today. The data is trapped in PDFs. This weekend, you're going to set it free.

Phase 1 — Crack the PDF

One memo. Nine checks. Prove it works.

We give you one real credit memo — a 22-page PDF for a $514,500 multifamily loan in Philadelphia. It has form fields, financial tables, analyst commentary, infographics, and property photos. Every content type that makes PDF extraction hard.

Your job: extract structured data from it, verify the math, and show where every number came from. Use any language, any framework, any AI. The only thing that matters is whether your output passes the scorecard.

Bronze
3 of 9 checks passing. "I can extract numbers and the math adds up."
Silver
6 of 9. "I can extract tables and cross-reference across pages."
Gold
All 9. "Every check passes." Full scorecard satisfied.
Phase 2 — The Loan File

Multiple documents. Cross-document verification. Build the pipeline.

A real loan file isn't one PDF — it's a folder. Credit memos, personal financial statements, rent rolls, appraisals, scanned tax returns. Phase 2 gives you the full picture and asks: can your pipeline handle all of it?

Your system classifies each document, routes it to the right extractor, and verifies data across documents — not just within one. This is what it takes to scale to 2,000+ memos.

Personal Financial Statement

Net worth, liquidity, assets and liabilities. Does it match what the credit memo says?

Rent Roll

Unit-level income for the property. Do the rents tie to the cash flow analysis?

Appraisal Summary

Property valuation and comparable sales. Does the appraised value support the LTV?

Scanned Documents

Tax returns, insurance certs, environmental reports. OCR quality varies. Can your pipeline handle it?

Cross-document verification checks

  • Borrower net worth on PFS matches the credit memo citation
  • Appraised value supports the LTV ratio in the loan terms
  • Rent roll income ties to the cash flow projections
  • Guarantor data is consistent across all documents
Pipeline
Auto-classify 3+ document types and extract structured data from each.
Crosscheck
Pass cross-document verification — data reconciles between documents.
Platinum
Handle scans via OCR, origination vs. servicing timeline, full loan file.

Start Building

Hackathon guide, team formation, schedule, awards, and the starter kit.

Join the Discord

Real-time chat, teams, help, and the Wall of Fails.

Clone the Repo

Starter kit, sample output, validator, schemas.

How it works

Clone and setup

git clone the repo. Run setup-check.sh. Look at sample-output.json to see what you're building toward.

Extract the PDF

Use any stack, any AI, any approach. The starter kit has a quickstart script that extracts your first field in 5 minutes.

Check your score

Run python validate.py your-output.json anytime. It tells you which checks pass and your current tier.

Watch the walkthrough
The business case and how the pieces fit together. Good context if you want the full picture.
Want the deeper story?
3 min

One-Page Brief

The problem, the pattern, the before-and-after. Why this matters to a bank.

Executive audience Read
12 min

Strategic Brief

The full business case. Why this data matters to a bank and what it's worth when it's unlocked.

Mixed audience Read
15 min

The Full Spec

The complete scorecard, output schema, reference architecture, and timeline. For teams who want every detail.

Engineering audience Read
Source files

Repo and Markdown sources

Everything is on GitHub. The Markdown versions render natively and are easy to copy into your own project.