What Intelligent Skill Extraction is
An ongoing research project on automated resume analysis. The system extracts structured skill data from unstructured resume text and computes a quantified compatibility score between a job description and a candidate profile. The research question is not simply which model is most accurate, but which model is accurate enough to be worth its inference cost in production.
My role: Researcher — pipeline design, model comparison, evaluation.
The problem it solves
Resume screening is keyword matching wearing a lab coat. A candidate who writes "built REST services in Express" is invisible to a filter looking for "Node.js", and a candidate who lists every technology they have ever seen ranks first. Extracting the underlying skills — and comparing them semantically to what a role actually requires — is the part that keyword search cannot do.
How it works
The work compares two families of approach on the same corpus. Transformer models (BERT variants) capture semantic equivalence between differently-worded skills but carry meaningful inference latency and hardware cost. Traditional ML pipelines built on engineered features and scikit-learn classifiers are dramatically cheaper to run but need far more hand-tuning to reach comparable recall. Both are scored on extraction accuracy and on inference cost per resume, because a model that cannot be served affordably is a paper result rather than a product.
The outcome
Ongoing. The framing carries over from my production work: correctness is a constraint, but the deployable answer is the one where the accuracy is worth what it costs to serve.
Technology used
- Python
- NLP
- Machine Learning
- BERT
- scikit-learn
Engineering highlights
- NLP-based skill extraction from unstructured resume text
- Automated job-to-candidate compatibility scoring
- Comparative study of BERT-based models vs. traditional ML pipelines
- Focus on production-viable accuracy/latency trade-offs