CleanControlling – AI-Based PDF Comparison Pipeline
An automated document monitoring and change-detection system that continuously compares regulatory PDF documents, identifies meaningful updates using deterministic diffing and AI-assisted semantic analysis, and proactively notifies stakeholders. The project prioritizes reliability, auditability, and real-world compliance workflows over heuristic or black-box analysis.
Role
Software Engineer (End-to-End Ownership)
Team
Solo project in direct collaboration with business stakeholders
Duration
~3 months (design, implementation, production rollout)
Impact & Metrics
The system replaced a manual, error-prone review process with a fully automated, continuous monitoring pipeline, delivering measurable operational and compliance impact:
- ~100% coverage of monitored vendor documents
- Daily automated checks, eliminating irregular manual reviews
- Minutes instead of hours to detect and communicate relevant changes
- Reduced compliance risk by ensuring stakeholders always work with the latest approved documentation
- Zero ongoing maintenance overhead after rollout due to containerized deployment
The solution scaled linearly with the number of documents without increasing human workload.
Brief
Regulatory and safety documentation changes frequently, yet many organizations still rely on manual checks or infrequent reviews, introducing compliance risks and unnecessary operational overhead.
In this project, I designed and implemented a fully automated pipeline that continuously monitors vendor-provided safety data sheets, detects document changes at both structural and semantic levels, and distributes actionable updates to relevant stakeholders. The system replaces reactive, manual processes with a proactive, auditable, and scalable solution tailored for real compliance workflows.
Technical Approach
The architecture was intentionally designed to prioritize determinism, traceability, and auditability before introducing AI-driven interpretation.
Document retrieval & versioning
Vendor PDFs are periodically fetched and compared against locally versioned documents.
PDF-level diffing
Structural and textual differences are analyzed to distinguish meaningful content changes from cosmetic updates (e.g., formatting or metadata changes).
AI-assisted semantic analysis
When changes are detected, a large language model (Gemini) generates concise, human-readable summaries highlighting what changed and why it matters, optimized for non-technical stakeholders.
Automation & delivery
The pipeline runs in a containerized environment (Docker / Docker Compose) and automatically distributes updated documents and summaries via email.
This hybrid approach avoids black-box decisions while still benefiting from AI where it meaningfully reduces cognitive load.
Learning and Growth
This project significantly deepened my experience in production-grade automation and applied AI:
- Designing systems where correctness and explainability are more important than raw model output
- Integrating LLMs into deterministic pipelines without sacrificing trust or auditability
- Building maintainable, containerized services aligned with real operational constraints
- Translating regulatory and compliance requirements into concrete technical guarantees
It reinforced the importance of engineering judgment when deciding where AI should — and should not — be used.
Takeaway
Reliable software systems are not defined by how intelligent they appear, but by how predictable, explainable, and useful they are in real workflows.
This project demonstrates how AI is most effective when used as a supporting layer on top of well-structured, deterministic systems — enhancing human decision-making rather than replacing it.