Inspiration
TRACE-AIR started with a simple problem: aircraft parts can look perfectly legitimate on paper while the documentation behind them is not. The AOG Technics case made that risk tangible, with more than 60,000 engine parts sold with falsified certificates between 2019 and 2023. What caught our attention was that the issue was ultimately uncovered through document verification. We asked ourselves: could AI help quality teams notice suspicious patterns earlier, before a part is installed?
What it does
TRACE-AIR is designed for maintenance shops receiving used aircraft components. The user uploads certificates, invoices and historical records, and the system extracts the relevant information from documents with different formats.
It then combines rule-based checks with machine-learning anomaly detection to identify things worth investigating: inconsistent serial numbers, unusual supplier-part combinations, suspicious timelines or document structures that differ from previously verified files.
TRACE-AIR does not decide whether a part is safe or fraudulent. Instead, it highlights the exact information that deserves attention so the quality controller can investigate it and, when necessary, confirm the document directly with its issuer.
How we built it
We designed TRACE-AIR around two main AI components.
First, a pre-trained document-understanding model extracts structured information from heterogeneous scans and certificates. Second, an anomaly-detection model, such as Isolation Forest, can learn patterns from previously verified records and identify new files that differ significantly from those patterns.
We deliberately kept deterministic rules for problems that do not require AI, such as comparing two serial numbers or detecting incompatible dates. This lets machine learning focus on the problems where it adds the most value: understanding varied documents and finding less obvious patterns across historical data.
Challenges we faced
One of the biggest challenges was avoiding the temptation to treat AI as a magic fraud detector. An unusual document is not automatically fraudulent, and a well-made fraudulent document may look completely normal.
Another challenge is data. Aviation maintenance records can be confidential, heterogeneous and difficult to obtain in large labelled datasets. For that reason, our approach starts with pre-trained models and rule-based verification, and only introduces anomaly detection when enough verified historical data is available.
We also discovered that competitors already provide document-analysis tools for aviation. That forced us to narrow our positioning toward small independent maintenance teams, a lightweight receiving workflow and learning from anomalies across successive deliveries rather than simply digitizing documents.
What we learned
The biggest lesson was that good AI products are not always the ones that automate the final decision. In a safety-critical environment like aviation, AI can be more useful as a second pair of eyes: surfacing inconsistencies, showing its sources and helping a trained professional make a faster, better-informed decision.
We also learned that building the ML model is only part of the problem. Data quality, explainability, integration into existing workflows and measuring false alerts are just as important as model performance.
What's next
Our next step is to test TRACE-AIR with three maintenance shops over 90 days, comparing the normal review process with the TRACE-AIR-assisted workflow. We would measure review time, missed anomalies and false alerts before deciding how much value the ML layer actually adds.
Log in or sign up for Devpost to join the conversation.