AI HR Assistant & Automated ATS Platform
An HR platform that brings policy questions, leave workflows, recruiting, and candidate evaluation into one system. Specialist agents work behind a shared supervisor, with hybrid retrieval for grounded answers.
A team of agents. A coherent workflow.
- Supervisor routing with specialist agents
- BM25 + vector search + reranking
- Transactional outbox for background jobs
The problem and its constraints.
HR questions and recruiting tasks need both flexible language understanding and predictable workflows. This platform combines policy Q&A, leave management, applicant tracking, and candidate evaluation behind a shared agent supervisor.
- Policy answers need relevant, grounded evidence.
- Leave and recruitment actions need explicit control around model output.
- Document ingestion and background jobs must handle retries and competing workers.
What the evidence supports.
The CV describes Shaswot's work on supervisor routing, hybrid retrieval, ATS screening, and background processing. The public repository is a multi-contributor project; the architecture shown here describes the shared system, not a claim of sole authorship.
Inside the approach.
Search in more than one way
BM25 supplies keyword matches while pgvector supplies semantic matches. Rank fusion combines the candidates before a cross-encoder reranks them. The pipeline brings exact policy terms and broader meaning into the same retrieval process.
Separate responsibilities
A LangGraph supervisor routes requests to policy, leave, recruitment, clarification, and recap agents. Redis confirmation locks and explicit workflows provide controls around actions that should not be left to an unconstrained chat response.
Move durable work out of the request
The CV describes a PostgreSQL transactional outbox and workers using FOR UPDATE SKIP LOCKED. MinIO handles document storage; background jobs support ingestion and processing outside a single request's lifetime.
Results with context.
Automated tests
Test-suite count reported in the CV; not a fresh test run.
Pydantic validation, request handling, and observable services accompany the agent workflow. The CV reports OpenTelemetry/Arize Phoenix tracing and more than 735 automated tests. This is a reported project snapshot, not a live test result.
What this project demonstrates.
The project illustrates how retrieval quality, specialist routing, and ordinary distributed-system controls can work together around an LLM.
A useful next evaluation would compare retrieval and ATS failure cases across a fixed, documented test set. This is a suggested direction, not a claim of implemented work.