Covenant Sentinel

The Risk Dashboard
The Input

💡 Inspiration

The LMA asked: "How can Borrowers and Lenders efficiently gather and distribute relevant information to ensure compliance?"

Currently, "Keeping Loans on Track" is a manual nightmare. While banks use sophisticated algorithms to decide on a loan, the actual monitoring of that loan relies on manual labor. Junior analysts spend 20-30% of their week manually reading 100-page PDF agreements and typing financial ratios (Covenants) into Excel.

This manual process is:

Slow: It takes hours to onboard a new deal.
Risky: Human error in data entry leads to missed defaults.
Unscalable: Banks cannot easily monitor thousands of SME loans.

We built Sentinel to be the "Digital Spine" of commercial lending, turning unstructured legal PDFs into structured, live risk dashboards in seconds.

💻 What it does

Sentinel is a secure, desktop-native application that acts as a risk engine for Credit Teams.

Drag-and-Drop Ingestion: Bankers simply drop a standard LMA (Loan Market Association) facility agreement into the app.
Hybrid AI Extraction: The system uses Google Gemini 1.5 Flash to parse the legal text. It doesn't just "find numbers"—it understands complex conditional logic (e.g., "Leverage must not exceed 4.00x, tightening to 3.00x after 2027").
Instant Compliance Matrix: It generates a "Pass/Fail" dashboard for every financial covenant found in the document.
Commercial Export: With one click, the data is exported to a formatted Excel Risk Report, ready for Credit Committee review.

⚙️ How we built it

We architected the system to be Enterprise-Ready and Model-Agnostic.

Frontend: Built with React, TypeScript, and Tailwind CSS wrapped in Electron. This ensures the app runs locally on the banker's desktop, a critical security requirement for financial institutions.
Backend: A Python/FastAPI server that handles file processing and orchestration.
The Intelligence: We integrated Google Gemini 1.5 Flash for its massive context window (1 Million tokens), allowing us to process entire loan agreements without chopping them into fragmented chunks.
Data Pipeline: We used pdfplumber for text extraction and SheetJS to generate industry-standard Excel reports directly in the client.

🚧 Challenges we ran into

The biggest challenge was "Legal Logic vs. Simple Extraction." Early versions of the AI would find the number "4.00x" but miss the context (e.g., "Step down to 3.50x"). We solved this by engineering a robust "Chain of Thought" prompt that forces the model to extract the condition and the timing of the covenant, not just the value.

🏆 Accomplishments that we're proud of

Accuracy: The system correctly identified the "Step-Down" leverage ratio in our test LMA documents, a nuance that most regex-based tools miss.
Speed: Reduced the "Time-to-Spreadsheet" from ~2 hours (manual) to under 15 seconds.
UX Design: We built a "Dark Mode" interface that mimics the aesthetic of Bloomberg/Refinitiv terminals, making it feel native to a trader's workflow.

🚀 What's next for Sentinel

Private LLM Hosting: Allowing banks to swap Gemini for a locally hosted Llama 3 model for air-gapped security.
Historical Analysis: Uploading quarterly compliance certificates to track covenant performance over time.
Chat-with-Doc: Adding a sidebar to ask specific legal questions (e.g., "What is the cure period for a breach?").