JobFlow AI

Highlights

• ★Model beats traditional ATS on candidate–job relevance
• ★RAG over personal career corpus enables genuinely tailored resumes

Problem

Traditional ATS filters rely on lexical keyword matching. Qualified candidates get rejected on surface-level mismatches, and recruiters waste cycles on irrelevant shortlists. The CV-JD research (Nov 2024) proved a DeBERTa-v3 baseline could outperform ATS on relevance — the product layer turns that into a usable platform.

Approach

• —DeBERTa-v3-based semantic matching model trained to rank CV–JD pairs on true relevance (beats TF-IDF baselines used in most ATS engines)
• —RAG retrieval over a user's resume corpus + career history to auto-tailor each application to the specific job description
• —Six product pillars: profile store, aggregated job feed, AI tailoring engine, one-click apply, application tracker dashboard, AI-assisted interview prep
• —End-to-end reproducible ML pipeline with orchestration — designed for continuous evaluation as the model evolves

Stack

Outcome

• —Underlying CV-JD matcher published Nov 2024 as a research artefact
• —Platform layer in active development — web app shell, job aggregation, and tailoring engine scaffolded

Tags