I got tired of manually searching for jobs and writing cover letters for every application. So I built a tool that does it — reads my CV, finds relevant listings, and drafts personalized emails. 100% local. No paid APIs. No data leaving my machine.
Most job search tools are generic. They surface hundreds of listings that technically match your keywords but don't actually fit your profile. Then you spend hours reading each one, writing slightly different cover letters, and still not knowing which ones are worth applying to.
I wanted something that reads my actual CV, understands what I do, and only surfaces listings that genuinely match — then helps me write the application.
Reads your CV and uses it as context for all ranking and generation. Not keyword matching — semantic understanding.
Searches 3 Swiss job boards simultaneously. Results are deduplicated and ranked by relevance.
Two-stage ranking: fast embeddings for retrieval, then LLM reranker for precision. Best of both worlds.
Drafts personalized application emails grounded in both your CV and the specific job listing.
3-layer system to prevent the model from inventing credentials or misrepresenting your profile.
FastAPI backend with a clean web interface. Results stream in real-time as the agent works.
Fully local — everything runs on your machine via Ollama. No OpenAI, no Anthropic, no cloud calls.
Building a production-grade RAG pipeline is harder than tutorials make it look. The retrieval step is easy. The reranking, hallucination prevention, and making the whole thing feel coherent is where most of the work went.
Writing 142 tests for an AI agent is also an interesting challenge — you can't just assert exact outputs. I had to think about what properties the outputs should have, not what the exact text should be.
The 9 Architecture Decision Records document why I made specific technical choices — from the choice of embedding model to how I structured the hallucination defense. It's a habit I picked up from working on real projects and it's genuinely useful.