06 · Live demo
CareerAgent
AI agents for the job hunt — RAG chat over your CV with citations, live company research, CV tailoring, Kanban tracker.
CareerAgent chats over a CV with page-level citations and refuses questions the documents cannot support. A research agent runs a tool loop over live web search, then structured output (JSON mode, Pydantic, one self-correction retry) shapes the brief. The tailor agent retrieves CV evidence for a job description; every bullet carries the supporting quote, plus honest gaps.
There is no agent framework on purpose. Tool loop, structured output, and provider fallback (Groq, then Mistral) are a small amount of plain Python. Every LLM request logs tokens, latency, and list-price cost. An LLM-as-judge suite includes a hallucination trap where only a refusal counts as a pass.
The public demo runs on Hugging Face Spaces and Cloudflare Pages. The same container has Kubernetes manifests with probes, limits, and an autoscaler. API docs are disabled in production; CORS is pinned; LLM endpoints are rate-limited.
In the repo
- Cited RAG chat with refusal when the answer is not in the docs
- Research agent: tool loop + structured brief
- Tailor agent anchored to CV evidence
- Per-request cost metering
- LLM-as-judge evals, including a trap question