2025 · Completed
ALDEVI: Compliance Through AI
My master's thesis at DTU and Alfa Laval: a contract analysis system using LLMs and retrieval-augmented generation to detect deviations faster and more consistently.
Overview
This master's thesis was completed for the MSc in Human-Centered Artificial Intelligence at DTU in collaboration with Alfa Laval. The work focused on a practical industrial bottleneck: reviewing complex customer tender documents against internal standards, expert knowledge, and historical project material.
The result was ALDEVI, a proof-of-concept system for deviation detection built around a structured 'Company Truths' knowledge framework, prompt-program optimization, hybrid retrieval, and tuned text and vision embeddings. Across the final validation, the selected architecture achieved 88.4% recall with a 0.73 F1 score on new projects and demonstrated a path to nearly four times higher review throughput at a cost of under 5 EUR per project.
Because the implementation was developed inside a company collaboration, the source code and proprietary data cannot be shared publicly. Instead of a public repository, this entry links to the written thesis, which documents the methodology, anonymized evaluation setup, system architecture, and business impact analysis in detail.
Highlights
- Designed and validated ALDEVI, an end-to-end RAG system for contract deviation detection in collaboration with DTU and Alfa Laval.
- Reached 88.4% recall and a 0.73 F1 score on new projects while reducing manual review time from 30 to 8 hours per project.
- Presented on 7 August 2025 and awarded the top Danish grade of 12; the code remains private because it belongs to the company.