If I decided to keep pushing the project further, these are the improvements I would make next.

Add a human review path for ambiguous cases

The biggest gap right now is not basic classification. It is ambiguity around downstream job-record matching.

The next useful step would be a small review path for cases where the system is not confident enough to create or update a record safely.

Improve observability

The current logs and Telegram summaries are enough for a small personal workflow, but they are not enough for a broader operational setup.

If I kept working on this, I would add better structured visibility around:

• failed resolver decisions
• skipped downstream actions
• repeated failure patterns
• per-run operational outcomes over time

Expand the evaluation datasets

The current eval workflow is practical and useful, but it can still be broadened.

The next step would be to add more edge cases, especially around:

• same-company multi-role ambiguity
• generic updates with weak role signals
• recruiter phrasing that hides the actual role title

Add a lightweight review-friendly output for job-record decisions

Right now the service applies downstream job-record decisions directly once they are resolved.

A helpful next step would be a small reviewable artifact for those decisions so they can be checked more easily when needed, without going through raw logs.