Portfolio
AI Agents & Pipelines

Autonomous Credit Covenant & Compliance Agent

Enterprise-grade AI agent for automated verification of corporate credit loan agreements — 93.3% exact match accuracy, 1.65ms BM25 retrieval, 86.9% token reduction on 50-page legal contracts.

GitHub Repository
93.3%
Accuracy
Exact match vs. official ground truth down to the cent
1.65ms
BM25 Retrieval
Average search latency per document
86.9%
Token Reduction
Input token savings on 50-page legal contracts
0.00%
Math Errors
Sandboxed Python arithmetic, zero LLM math
The Problem

Financial institutions spend thousands of hours manually auditing loan dossiers to verify covenant compliance (Leverage Ratios, Capex Limits, Related-Party Transaction Caps). Human error rates are high, and each audit cycle takes days.

The Solution

A modular, decoupled pipeline: multithreaded PDF ingestion with OCR fallback → deterministic regex + LLM classifier → Okapi BM25 paragraph ranker → sandboxed Python arithmetic engine (zero LLM math) → multi-currency FX calculation → automated HTML compliance report & executive dashboard.

Architecture Highlights

01

Zero LLM arithmetic — sandboxed Python float engine handles all calculations

02

86.9% input token reduction via BM25 paragraph snippet retrieval on dense legal PDFs

03

Multi-currency FX engine (KZT, EUR, RUB → USD) with dynamic rate extraction

04

Bi-directional marginal transaction selector — identifies single transactions that flip breach status

05

100% offline execution capability via Ollama on Apple Silicon

06

Auto-generated HTML executive dashboard + machine-readable audit trail JSON

Tech Stack

Python 3.11Ollama (Apple Metal GPU)Okapi BM25PyMuPDF / pdfplumberTesseract OCRPydantic v2DockerClaude 3.5 Sonnet

Want something similar?

Available for Founding Engineer roles & architecture contracts.

Discuss ProjectAll Projects