Job Tracker AI

The problem

Application updates get buried in Gmail and Outlook. Automatically tracking them requires more than finding keywords: the app must distinguish hiring events, extract stated facts, and avoid changing the wrong application.

What I built

• GPT-6 Luna for classification. OpenAI Decisions checks whether an email concerns a specific application, then classifies its event as an application, assessment, interview, offer, rejection, or other update.
• GPT-5.4 mini for extraction. Responses with Structured Outputs returns company, role, requisition, explicit dates, and a source excerpt in a strict Pydantic-derived schema.
• A validation and review layer. Python checks evidence, identity, and chronology before updating SQLite. Ambiguous proposals go to human review; missing facts remain empty.
• A usable desktop release. A PySide6 app with Gmail and Outlook imports, application history, background workers, resumable scans, cached inference, and user-selected spending limits.

Why I built it this way

• Separate classification from extraction. Relevance and event type are bounded Decisions questions; extracted fields need a custom schema. Separating the tasks avoids extraction for clearly unrelated mail and makes each stage easier to inspect.
• Treat valid JSON as a starting point. A schema checks structure. Evidence checks, record matching, and chronological rules decide whether the proposed update belongs in the application history. Neither model writes to the database.
• Keep uncertainty visible. Low confidence, refusals, missing evidence, or unclear identities lead to review. Email is untrusted input; the models have no tools for following links, executing code, or modifying the mailbox.
• Reserve cost before dispatch. SQLite records spending reservations and scan progress before model calls. Completed outputs can be reused; interrupted scans resume explicitly. A billed request whose response was lost can still cost money again.

What I learned

I practiced prompt engineering, model routing, Structured Outputs, and human-in-the-loop design. The useful engineering happens after a response arrives: validating evidence, handling uncertainty, preserving corrections, and making cost and recovery understandable to the user.