Grounded Job Agent
An evidence-backed job-fit and CV preparation pipeline using a synthetic candidate. Draft bullets cite evidence IDs and code checks recognized technology terms. These are bounded checks, not a general truth detector. A person reviews each package; the agent never submits applications.
1 · RetrieveLocal lexical retrieval selects evidence for each requirement.
2 · AssessEach requirement is met, partial or a gap, with cited evidence. Hard gaps cap the tier.
3 · TailorBullets cite evidence IDs; unsupported recognized technologies block the package.
4 · GatePDF checks bound to the file hash, then human approval.
Example runs
What happens to an invented line
This hand-authored negative example adds Kafka and an uncited responsibility. It shows the checker responding to those inputs, not a measured LLM failure rate: