shabs-ch/automotive-reskill

🚗 Automotive Reskill

AI-powered career transition tool for German automotive engineers moving into AI roles.

Live Demo

[Try it here →](https://huggingface.co/spaces/shabs-ch/automotive-reskill) | [GitHub →](https://github.com/shabs-ch/automotive-reskill)

Paste your CV, get matched AI roles, understand your skill gaps, and receive a personalised learning roadmap with real, cited courses — including Bildungsgutschein-eligible options.

Why I built this

The German automotive industry is under significant disruption. Engineers with 10-20 years of experience in embedded SW, testing, systems engineering, and program management are facing displacement — but their skills transfer strongly to AI roles that most tools and recruiters can't see.

I built this to make those invisible skills visible, and to give engineers a concrete, personalised path forward.

What it does

Paste CV
↓
Extract skills + classify profile into AI role families
↓
Match against 60 curated job ads (semantic search, ChromaDB)
↓
Analyse skill gaps for chosen role (ontology + mappings + Claude)
↓
Generate month-by-month learning roadmap (RAG, cited courses only)
↓
Show Bildungsgutschein-eligible courses + application guidance

Architecture

src/skill_extractor.py     → CV → structured skills JSON (Claude)
src/profile_classifier.py  → classify profile into 4 role families (Claude)
src/role_matcher.py        → semantic search against job corpus (ChromaDB)
src/gap_analyzer.py        → gap report: have / transfers / need (Claude)
src/roadmap_generator.py   → RAG roadmap citing real courses (Claude)
src/bg_checker.py          → Bildungsgutschein guidance (data lookup)
src/agent.py               → fixed-sequence orchestrator, state management
src/chroma_client.py       → auto-detects local vs ChromaDB Cloud

Key design decisions

Classifier before retrieval — semantic search alone fails for management profiles. A Claude-based classifier routes profiles to the right role families before ChromaDB retrieval.

RAG for roadmap grounding — the roadmap generator retrieves courses from a 108-course corpus and passes them to Claude. Claude can only cite courses that actually exist — hallucination rate is zero on test cases.

Weighted course scoring — course quality (provider reputation) and skill match take priority over Bildungsgutschein eligibility. A €49 Coursera course from Andrew Ng outranks a mediocre free German course.

Static YAML over vector DB for knowledge — skill ontology, automotive→AI mappings, and course corpus are stored as YAML files, not embedded in ChromaDB. Rationale: these are structured, tagged datasets where keyword + metadata filtering is more precise than semantic similarity. YAML is human-readable, version-controllable, and directly inspectable. ChromaDB is used only for job ads where semantic similarity search genuinely adds value.

Prompted Claude over fine-tuning — evaluated QLoRA fine-tuning for CV reframing. Chose prompted Claude with few-shot examples instead. Insufficient training data (<50 examples), and prompted approach produces equivalent quality at this scale. Would revisit at >500 labelled examples.

Static corpus for eval reproducibility — 60 curated job ads and 108 courses kept static for v1. Live API integration (Bundesagentur für Arbeit) is a documented v2 priority.

Evaluation

Built a 15-case hand-graded eval framework across 6 engineer archetypes. Temperature=0 for deterministic results.

Retrieval quality was evaluated qualitatively — retrieved courses consistently matched identified skill gaps across all 5 integration test scenarios. Formal IR metrics (Recall@K, MRR) are a v2 priority.

User Research

• —6 user interviews conducted before building (Week 1)
• —Profiles: PMO Lead, Systems Requirements Lead, Technical PM, SW Requirements Lead, SW Test/Validation Lead, SW Project Manager
• —Top pain: "I can't see how my automotive skills transfer to AI roles"
• —Key finding: 4/6 prefer automotive-adjacent transition over full domain switch — shaped the domain preference filter

Stack

Setup (local)

git clone https://github.com/shabs-ch/automotive-reskill
cd automotive-reskill
python -m venv .venv
source .venv/Scripts/activate  # Windows Git Bash
pip install -r requirements.txt

Create .env:

ANTHROPIC_API_KEY=your_key
CHROMA_API_KEY=your_chroma_key      # optional — uses local ChromaDB if not set
CHROMA_TENANT=your_tenant_id
CHROMA_DATABASE=automotive-reskill

Never commit your .env file. It is gitignored by default.

# Embed job ads into ChromaDB (first time only)
python src/embed_jobs.py

# Run the app
streamlit run app.py

Limitations and future work

v1 is deliberately scoped. Known limitations and v2 priorities:

• —Job corpus — 60 static curated ads. v2: live Bundesagentur für Arbeit API integration
• —BG eligibility — 23 confirmed, 16 unknown. v2: AZAV database cross-reference
• —Retrieval metrics — quality validated qualitatively. v2: Recall@K, MRR measurement
• —Language — English UI only. v2: German localisation
• —No user accounts — session state only, no persistence across visits
• —CV upload — text paste only. v2: PDF upload via document parsing
• —Mobile experience is limited — optimised for desktop use

These improvements are well-defined and implementable — deferred by choice to broaden AI engineering exposure across multiple projects.

Project notes

All architectural decisions, engineering learnings, and product decisions are documented in:

notes/decisions.md      → every non-obvious architectural choice
notes/learnings.md      → running engineering learnings log
notes/feedback_round1.md → user feedback and fixes
notes/future.md         → v2 ideas and backlog
docs/architecture.md   → detailed component architecture + design decisions

License

MIT