HeadHunter Agent — German IT Job Application Assistant

📦 Base model: llama3.2:3b-instruct-q4_K_M
🖥️ Optimized for: CPU-only inference (no GPU required)
🇩🇪 Primary language: German (documents) + Russian (analysis reports)
🔗 Project: HeadHunter PROJECT_beta

What This Model Does

HeadHunter Agent is a role-conditioned LLM that automates the German job application process in 5 structured steps:

ATS Scoring Formula (Textkernel-weighted)

ATS Score = (Job title match × 30) + (Tier1 skills% × 0.45) + (Tier2 skills% × 0.15) + 10

• ≥ 70 → Apply now
• 50–69 → Adapt CV first
• < 50 → Skip vacancy

Who Is This Model For

✅ IT job seekers applying to positions in Germany (2026 market)
✅ Career coaches and HR consultants working with German ATS systems
✅ Developers building job application automation pipelines
✅ Students and bootcamp graduates entering the German IT market
✅ Non-native German speakers who need DIN 5008-compliant documents

Key Features

• 🎯 ATS-aware generation — knows Workday, Personio, SAP SuccessFactors, Greenhouse, Softgarden/onlyfy, Lever, d.vinci, Recruitee
• 📋 Structured JSON output for Step 1 and Step 2 (machine-readable, pipeline-friendly)
• 🚫 AI-language blacklist — avoids generic AI phrases: dynamisch, leidenschaftlich, begeistert, vielfältig, umfangreich (common ATS red flags)
• ⏱️ Recency Boosting — weights recent experience (< 3 years) higher, following Textkernel ranking logic
• ⭐ S.T.A.R. bullet format — Situation + Task + Action (Technology) + Result (Number)
• 🏛️ EU AI Act compliant — enforces Human-in-the-Loop; model adapts, never invents facts
• 📄 DIN 5008 cover letter — correct German business letter format, no forbidden clichés
• 🔤 Softgarden naming rules — PDF filename must start with CV_[Name]_[Firma]_Lebenslauf.pdf

Hardware Requirements

Optimized parameters for CPU-only (no GPU):

num_gpu 0 | num_thread 4 | num_ctx 1024
num_batch 128 | mirostat 0 | temperature 0.10

Expected speed on Intel i5 (no GPU): ~4–8 tokens/sec
Expected speed on RTX 4070 (GPU): ~40–80 tokens/sec

Quick Start

# Pull and run:
ollama pull lelegioner/headhunter-public
ollama run lelegioner/headhunter-public

# Example prompt:
"Analysiere diese Stelle: IT-Systemadministrator bei Deutsche Telekom.
Anforderungen: Linux, Windows Server, Active Directory, Azure, Python."

Expected response (Step 1):

{
  "firma": "Deutsche Telekom",
  "stelle": "IT-Systemadministrator",
  "ats_system": "SAP SuccessFactors",
  "hard_skills_tier1": ["Linux", "Windows Server", "Active Directory", "Azure"],
  "hard_skills_tier2": ["Python", "Docker", "Monitoring"],
  ...
}

Customize for Your Profile

This public version contains placeholder fields instead of personal data.
To use it for your own job search, provide your profile in the first message:

MEIN PROFIL:
Name: [Your Name] | Email: [your@email.com] | Verfügbar: [Date]
Kenntnisse: [Your skills, comma-separated]
Erfahrung:
  [Date]-heute: [Role] | [Company] → [Technologies]
  [Date]-[Date]: [Role] | [Company] → [Technologies]

STELLENANZEIGE:
[Paste the full job posting here]

Limitations

❌ What This Model Is NOT For

• ❌ General-purpose chat — not optimized for conversation, coding help, or Q&A
• ❌ Non-IT professions — scoring formula calibrated for IT roles (Admin, DevOps, Dev)
• ❌ Job markets outside Germany — ATS logic, DIN norms, and Textkernel weights are DE-specific
• ❌ Fully automated submissions — always requires human review before sending (EU AI Act)
• ❌ Generating fake experience — strictly refuses to invent skills, dates, or employers
• ❌ English-language applications — cover letters and CVs are generated in German by default
• ❌ Legal or HR compliance advice — provides formatting guidance only, not legal counsel
• ❌ Real-time job search — cannot browse job boards or query APIs

Ethics & Compliance

• EU AI Act (Annex III): CV screening and candidate selection tools are classified as High-Risk AI systems. This model is a drafting assistant, not a decision-maker. Final submission requires explicit human approval.
• GDPR: Do not include third-party personal data in prompts. The model does not store or transmit data.
• Transparency: Employers are not notified that documents were AI-assisted. Users are responsible for disclosure per applicable law.
• No discrimination: Model does not make recommendations based on age, gender, nationality, or health status.

Full Project

The complete HeadHunter PROJECT_beta includes: - agent.py — full automation pipeline (5-step workflow, SQLite logging, PDF generation) - Modelfile.cpu / Modelfile.template.gpu — hardware-specific builds - ATS database, skills map DE 2026, EU AI Act compliance guide - Opencode CLI integration via MCP server

Candidate profile in this public version: replaced with placeholders — safe to share and adapt.

License

Base model: Meta Llama 3.2 Community License
System prompt and workflow logic: MIT
Use of this model for job application automation is permitted under both licenses. Commercial use requires registration with Meta per the Llama 3.2 Community License terms.