End-to-End Hiring Pipeline

Candidates submit resumes through branded career pages or ATS sync. PDFs are parsed in real-time by EdgeParse — a pure-Rust parser that extracts Markdown in 0.064 seconds per document with a 0.787 benchmark score. No JVM, no GPU required.

Groq LLM processes extracted Markdown through a two-stage pipeline: first structuring into canonical sections (education, experience, skills, projects), then extracting typed fields. Link extraction uses regex as ground truth with LLM fallback for ambiguous cases.

Four separate all-mpnet-base-v2 ONNX INT8 embeddings are generated for each resume: Skills, Projects, Experience, and Summary. Section-wise embedding outperforms whole-document embedding by 11.4pp on Precision@10 because each section operates in its own semantic space.

Embeddings are stored across four Qdrant collections accessed via gRPC (p50 latency 1.11ms vs 2.40ms REST). PostgreSQL stores metadata, extracted links, and Cloudinary PDF URLs. A single SQL query fetches all four vector IDs, reducing round trips from 4 to 1.

Three-stage retrieve-then-rerank: HNSW vector retrieval builds a 2K candidate pool (0.20 weight), cross-encoder re-ranking narrows to top-K (0.65 weight), and BM25 lexical scoring recovers exact-match candidates (0.15 weight). Cheapest operation runs first.

Shortlisted candidates enter the gated assessment pipeline. Stage 1: coding assessment with Judge0 CE and Tauri OS-level proctoring. Stage 2: AI avatar interview with offline Supertonic 3 TTS and adaptive LLM questioning. Stage 3: live recruiter session via LiveKit SFU.

All assessment signals are fused into a composite score with configurable weights per role. Recruiter receives a full evidence package: resume summary, coding scores, interview transcript, proctoring logs, and integrity confidence. Offer or reject with data-backed decisions.