Store Ticket Assistant

Summary

A local-first assistant for a workflow I know well: capturing support-call details into a structured ticket. The interesting engineering is in what it refuses to do (leave the machine) and in how transcript quality is handled without a cloud model: a hallucination classifier, a reversible correction dictionary, and regression suites built from real transcripts after unit tests missed two analyzer regressions.

The problem and who it serves

Support technicians re-type the same call details into tickets, and sending internal call notes to a cloud service is not acceptable. The tool had to capture a call once, privately, on the technician’s own machine.

Store-technology support technicians writing tickets from calls. Built on personal time and equipment with no employer data; names of ticketing products and internal fields are deliberately absent from this page.

My role

Solo. I set the privacy stance (everything local; any model call restricted to localhost), designed the transcript-quality pipeline, the release checklist and the regression suites. Implementation was AI-assisted under my direction.

Architecture

The boundary is explicit: Rust owns side effects, TypeScript owns pure logic, and the Tauri capability file is the privacy contract. The HTTP capability allow-lists exactly two localhost origins, so even the optional model path cannot reach the network.

Decisions and trade-offs

• Route model calls through the shell, not the webview. Browser fetches to a local model fail CORS preflight. Instead of asking users to reconfigure their model server, requests go through the Rust HTTP plugin with capability-scoped URLs.
• Rules before models for speaker turns. Diarisation models are large and need training data; a rule-based question/answer turn-taker with advisory labels and one-click relabel was good enough and fully testable.
• Regression suites from real transcripts. After unit tests passed while the analyzer regressed twice, two scripts replay recorded transcripts (12 transcripts, 73 field checks) in CI.
• Reversible corrections. Every correction rule records what it changed so a wrong fix can be excluded from forms rather than silently baked in.

Verified results

Limitations and remaining work

• Prototype, not in production use; accuracy figures in the README (filter and speaker-label rates) are self-reported and are not repeated here.
• Rust side has no tests of its own; the logic under test is the TypeScript layer.
• Two UI pages are placeholders; releases are manual per platform.