Application Pipeline

Listings are collected via headless browser automation, normalised, and stored in a local SQLite database. From there a pipeline view walks each role through tailoring, review and submission, and a tracker records what happened. Resume and cover-letter generation produce a real PDF, and approving a role generates both in one step.

It is a single-user local application by design. There is no account system, no hosted database and no server round trip for the data, because none of those would make it better and all of them would make it slower to use.

The tool routes work to different providers by task rather than settling on one. Long-context reasoning over a full job description and resume goes to Claude. High-volume, latency-sensitive classification — is this listing relevant, is it a duplicate, does it match the criteria — goes to Groq, where inference is fast and cheap enough to run over hundreds of listings without thinking about it. Gemini handles a third class of extraction work.

The point of the abstraction is not provider-agnosticism as a virtue. It is that the cost and latency profile of 'read this carefully and write something good' is completely different from 'sort these five hundred things', and paying frontier prices for the second is how a tool like this stops being worth running.

The tailoring step originally had free rein over the resume. It produced better prose and, occasionally, quietly better job titles — an 'SDE 2' became a 'Senior Software Engineer', a scope grew slightly in the retelling. Every individual edit was defensible. Collectively they were a resume that misrepresented me, generated by a tool I built to help me.

So certain fields became structurally immutable: titles, dates, employers and metrics are pinned and the model cannot touch them regardless of what it is asked to do. It can select, order and phrase. It cannot change facts.

That is the transferable lesson, and it is the same one as the server-side cart validation two case studies up. A generative step is a system boundary. You validate at boundaries — you do not rely on the thing on the other side behaving well, and 'the prompt says not to' is not a constraint.