Projects / P-10
Data, code & AI
Signal detection agent with a SQL database
Extending the catheter analysis into an agent that codes reports, checks its own answers, stores everything in a database and runs SQL checks for emerging problems.
P-102026
- Project
- An auditable AI agent that flags emerging device problems
- Context
- Self-directed
19serious events flagged for review
11automated tests
SQLitedatabase with full audit log
The question
Can the coding method be turned into a repeatable system that catches problems early and keeps a person in charge of the serious calls?
What I did
- Built an agent that codes each report, checks the result against the coding rules, retries with the errors fed back, and sends serious or uncertain cases to a human review queue.
- Designed a SQLite database for reports, labels, the review queue, signals and a full audit log.
- Wrote SQL trend checks with views, CTEs and window functions: rising failure modes, keyword clusters and serious events.
What I found
In a run on the 300 coded reports, the agent flagged the 2023 septum leak and air-entry clusters, rises in kinking and packaging defects in 2025, and 19 serious events, then drafted a signal report for sign-off. A simple 'doubling' rule missed the gradual rise in catheter breaks, so I added a share-based check, kept separate from the pre-set rules.