Inside the Build / 03
AutoHunt: from CV to reviewed application
Follow a job application through a stateful AI workflow, with a human decision between preparing the material and submitting it.
The task
Start with a resume, discover relevant roles, and prepare an application the candidate can review. This walkthrough summarizes the architecture documented in the project overview; the execution sketches are illustrative pseudocode.
- PREPARE · STAGES 01–06
- Parse → discover → match → draft → critique → refine
- REVIEW · STAGE 07
- Checkpoint and pause for the candidate’s decision
- ACT AND RECORD · STAGES 08–10
- Submit approved material → track the outcome → compose a digest
Through the workflow
Expand a stage to inspect its input, output, and purpose.
- INPUT
- A resume PDF
- OUTPUT
- Validated candidate data
pdfjs-dist extracts the resume text. AutoHunt turns it into structured professional data, including experience, education, skills, and contact information, with Zod schemas checking the expected shape.
text = extract_pdf(resume)
candidate = parse_professional_data(text)
validated = candidate_schema.parse(candidate)WHY THIS MATTERS
Later stages need consistent fields rather than repeatedly interpreting raw PDF text. Structured data also makes missing or malformed information easier to inspect.
What this reveals
The workflow coordinates both data and authority: structured candidate context informs matching and drafting, checkpoints preserve progress, and approval controls the transition to browser automation.
The result is a process the candidate can inspect before action and audit afterward, rather than a single prompt that tries to do everything at once.
More walkthroughs