PRAXA is an AI career assistant for people looking for a job. It finds job postings that suit you, writes a CV and cover letter for each one without making anything up, and lets you practise the interview out loud. It started as a tool for one engineer job-hunting in Germany and now works for anyone: you name the roles you’re aiming for, upload your CV, and PRAXA works from your real experience.
Everything it says about you comes from your own CV. Everything it knows about a field comes from a library of career guides and seven expert packs: automotive cybersecurity, DevOps, AI engineering, software engineering, data engineering, embedded software, and QA and test automation.
How it’s built
PRAXA has a Python backend and a React frontend, served together as one app.
The backend is Python 3.12 with FastAPI. It uses Supabase Postgres for data and sign-in, with row-level security so each account only ever sees its own rows, and Chroma as the vector database. The models run through OpenRouter, and every call goes through a single gateway that checks the budget first and records the cost.
Search mixes meaning-based and keyword search (fused with RRF) and reranks the results with a cross-encoder. A grader then judges whether the evidence is good enough and searches again if it’s thin.
The agent is my own tool-calling loop, not LangChain or LangGraph. Each chat mode gives the model a different set of tools, and the server refuses anything outside that set. The interview follow-ups and “Should I apply?” are plain rules, deliberately not model calls.
A small neural network trained in PyTorch sorts postings into the expert packs. Mistral checks for jailbreak attempts and personal data. LangSmith traces each answer, with personal data masked first. The spoken interview uses Google’s Gemini Live API with a one-use token, so the API key never leaves the server.
The frontend is React 19 with TypeScript, Vite and Tailwind CSS. It all runs on Google Cloud Run starts the daily agent each morning.
