Parselex: A Cascaded Small-Encoder Pipeline for Structured Resume Extraction

Problem. Resume parsing today is usually done by routing PDFs through a general-purpose LLM API — real costs even for a small, fixed-schema task: per-call fees, third-party transmission of personal data, non-deterministic output. Approach. Parselex is a resume parser built as a 13-stage cascade of small (~22M-parameter) encoder classifiers — a single MiniLM backbone fine-tuned per stage, combined with CRF sequence decoding and a gated spatial-fusion layer over PDF layout features — running entirely on CPU with no external API calls. Results. Validation accuracy across all 13 stages, a three-way comparison against a cloud LLM (Gemini) and a local LLM (Gemma), a latency benchmark, and per-stage model sizes. On flat field-value extraction, both LLM baselines beat the pipeline on accuracy (F1 0.848 and 0.790 vs. 0.669); Parselex's case rests instead on cost, latency, and privacy. Availability. Code, trained weights, and a runnable inference API are publicly available.