Amendment Radar for LMA

The syndicated loan market trades over $1 trillion annually, yet its operational backbone relies on a process from the 1990s: "Stare and Compare."

Comment

Inspiration

There's a $500 million deal sitting on a trading desk right now. The Ops team has 72 hours to book it. But they can't, because somewhere in a 200-page PDF is a change that will blow up their covenant calculations. They just don't know where.

Today, that takes 5 days of manual review. We do it in 60 seconds.

What it does

Amendment Radar is post-execution intelligence for the secondary loan market.

  • Ingests raw credit agreements + amendments
  • Filters out formatting changes and typo fixes
  • Highlights what actually matters: covenant tightening, transfer restrictions, reporting frequency spikes
  • Outputs trade-ready data: Obligation Ledger, Evidence Pack, Materiality Scores
  • The AI surfaces changes. The banker makes the call. That's the design.

How we built it

We prioritized the "Banker UX" over the plumbing. We built a high-fidelity React dashboard that looks and feels like production software. For the backend, we engineered three specific "Gold Standard" scenarios (Alpha, Beta, Gamma) with pre-computed data to demonstrate exactly how the "Materiality Scoring" should work in a perfect world. We did implement a Gemini fallback for custom text, but for this demo, we focused on the "Happy Path" to prove the commercial value.

Challenges

The biggest challenge was defining "Materiality". To an LLM, "quarterly" vs "monthly" is just a one-word change. To a banker, it's a massive operational burden. We realized that a generic "summarize this" prompt is useless. The system needs a strict scoring model (e.g., "Covenant Tightening = 100pts") to be trustable.

Accomplishments

  • 60 seconds to surface 8 material changes
  • Clause-level provenance on every extraction
  • UX that looks like production software, not a hackathon demo

What we learned

A working prototype is easy. A trustable financial tool requires obsessive attention to edge cases, what happens when the text is malformed, when the AI hallucinates a clause, when confidence is low. The hard work is making it fail gracefully.

What's next

Actually building the plumbing.

Hooking up a real local OCR pipeline to ingest raw PDFs (so we don't have to copy-paste text). connecting that "Obligation Ledger" JSON to a real API.

Built With

Share this project:

Updates