Skip to content

Profile overhaul Phase 4: import from resume, LinkedIn PDF and GitHub #136

Closed

Description

@yash10019coder

Phase 4 of epic #117 (Profile System Overhaul). Plan: docs/plans/2026-10-05-001-feat-profile-import-plan.md; rules contract: docs/plans/2026-10-05-002-profile-import-rules.md (every rule ID gets a test). Stacked on #135 → #133 → #125 (none merged yet).

Goal

Fill the profile from a resume, a LinkedIn PDF export or a GitHub username so the user stops typing contact details, links, location and skills, and so the system knows which skills belong to which role or project (years per skill are computed in code and shown, not yet used to answer questions). Nothing inferred is written without per-field/per-entry review, and nothing imported overwrites what the user set, learned or locked.

Decisions

  • Rule-based parser + review/diff UI + consent first; the LLM extractor is built but off until providers are listed in an env allowlist (empty = rule-based only, fail closed).
  • Imports profile fields (name, phone, employer, location, links, skills → target_tags, titles → target_titles, headline) and experience/project entries with their own skills (new ResumeEntry). Education is not stored.
  • Sources: resume + LinkedIn PDF through one pipeline, then GitHub (public API, no OAuth). No scraping.
  • stdlib re with hard input bounds (no new native dependency); values the user edits in review are written as user.

Units

U0 plan/rules/issue · U1 models · U2 write APIs + years-per-skill · U3 document safety · U4a import_eval command · U4 field rules · U4b entry rules · U5 tasks/service/sweep · U6 web (rule-based path, mergeable) · U7 move llm_providers · U8 LLM extractor + gate · U9 GitHub · U10 docs/statuses.

Measured on 81 real resume PDFs (aggregate only)

The first draft of the rules would have rejected 37% of real resumes (benign /OpenAction), missed 31% of names, and mislabelled 27 ordinary resumes as LinkedIn exports; the rules now encode the fixes. 21% of date ranges are in Education. Rules alone fully resolve ~56% of experience entries (24% title only, 20% ambiguous), which is why entries default to reject unless complete and why the LLM path matters later. No real LinkedIn export was available, so those rules are provisional.

Not included

Education storage; answering "years of X" questions from the computed years (follow-up); editing entries after import; #130 reworded-question matching; #129 (should land before the LLM path is switched on).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:auto-applyapps/auto_apply Greenhouse auto-apply pipelineenhancementNew feature or requestseverity:P1High-impact defect likely hit in normal usage

    Projects

    • Status
      Done

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions