HireMate – AI Assistant for Finding and Applying to Jobs

Features

• Searches LinkedIn jobs via Apify actors
• Tailors your CV to each role using OpenAI
• Emails you the job links with tailored CV attachments
• Optionally lists 1st-degree connections at target companies (from a CSV export)

Tools Used

• Apify Client: Fetches LinkedIn job results from your scheduled actor runs (no re-triggering).
• OpenAI: LLM for structured CV extraction, edit plan generation, JD topic and priority extraction, and deterministic rewrite to structured JSON.
• SMTP (email): Sends tailored resumes with job links as attachments.
• Pydantic: Models and settings validation.

Notes

• Requires an Apify token (APIFY_TOKEN) and an OpenAI API key for tailoring.
• Email uses SMTP; set DRY_RUN=true to preview without sending.
• Without OPENAI_API_KEY, tailoring falls back to your original resume text.
• Connections CSV is optional; export from LinkedIn and set LINKEDIN_CONNECTIONS_CSV.
• Duplicate protection: processed job links are stored in state/sent_jobs.json and skipped in future cycles.

Setup

1. Create and fill .env based on .env.example.
2. Install dependencies:

pip install -r requirements.txt

3. Run commands:

# Search-only to CSV (uses settings from .env for filters)
python get_jobs.py search --max-jobs 5 --out jobs.csv

# Full pipeline: search, tailor resumes, and email
python get_jobs.py pipeline --max-jobs 3

# Watch mode: poll every 60 minutes and process only new jobs
python get_jobs.py watch --minutes 60 --max-jobs 5

Logic Used

1. Resume Text Extraction

   • Detects file type (PDF/DOCX/TXT) and extracts raw text.

2. LLM-based Structured Extraction

   • Parses resume to a structured entity(name, contacts, summary, skills, experience/academic).
   • Results are cached on-disk keyed by resume hash.

3. Job Understanding

   • LLM extracts canonical JD topics and selects top priorities evidentially(!) supported by the resume and

4. Pass A – Evidence-backed Edit Plan

   • Enforces zero fabrication, evidence-backed resume edit plan.
   • Returns JSON with suggestions and missing job description unsupported_requirements.

5. Pass B – Deterministic Rewrite to Structured JSON

   • Applies ONLY the approved plan to the structured resume, producing a final structured payload for the template builder.

6. Template-based DOCX Generation

   • Renders contact, summary, skills, experience, academic items.
   • output DOCX.

7. Review Artifacts

   • For each job, a Company_Title.review.json is written to the output folder with gaps and unsupported_requirements for quick audit of missing evidence.

Design Decisions

Why Two-Pass Architecture (Pass A + Pass B)?

• Pass A (Plan): LLM generates evidence-backed edit plans with strict constraints. This prevents fabrication and ensures all changes cite resume text or EXTRA_KEYWORDS.
• Pass B (Apply): deterministic rule-based application of the plan to structured data. No LLM calls - just mechanical string replacement and array manipulation.
• Benefits: Eliminates scope creep, reduces costs (1 LLM call instead of 2), faster execution, and guaranteed consistency.

Why Sanitization (_sanitize_plan)?

• Model drift protection: LLMs sometimes return malformed JSON or unexpected field shapes. Sanitization normalizes these before processing.
• Safety guardrails: Prevents title overwrites, validates that "find" text exists before replacement, and gates additions to wanted/extra topics only.
• Deterministic safety: Ensures Pass B receives clean, validated input for reliable rule-based application.

Why Template-Based DOCX Generation?

• Consistency: Clean, predictable layout regardless of input resume format.
• Maintainability: No complex in-place editing or style preservation logic.
• Skills grouping: Easy to render organized skills with headings and comma-joined items.
• Alternative rejected: In-place DOCX editing was fragile and hard to debug.

Why LLM-Only Extraction?

• Accuracy: Heuristic parsing (regex, keyword matching) was unreliable for varied resume formats.
• Structured output: LLM provides consistent JSON schema for experience/academic items.
• Caching: Expensive LLM calls are cached by resume hash to avoid redundant API costs.

Why Evidence-Based Planning?

• Zero fabrication: Every addition must cite resume text or EXTRA_KEYWORDS.
• Audit trail: Gaps and unsupported requirements are logged for review.
• Quality control: Prevents the model from inventing experience or upgrading seniority.

Why Apify Over SerpAPI?

• Scheduled runs: Apify actors can run automatically every hour, eliminating API rate limits.
• Data freshness: Fetches latest results without triggering new scrapes.
• Cost efficiency: No per-query charges, just actor subscription.