RoleSense

An AI resume screener that classifies resumes into 40+ job categories and matches them to a job description — built for recruiters buried under applications (hi, if that’s you reading this).

Trail statsTelemetry · at a glance

~82%
accuracy classifying resumes (Random Forest)
43
job categories
5
classifiers trained and compared
3
resume formats accepted: PDF, DOCX, TXT

01 · The Map · Mission Brief

Brief

RoleSense is the tool on the other side of every resume submission: upload a batch of resumes (PDF, DOCX, or TXT) and a job description, and it scores each candidate two ways — a trained classifier sorts them into one of 40+ job categories, while a separate TF-IDF/cosine-similarity engine measures how closely their resume text actually matches the job description. A hybrid score blends both into a single pass/fail call with a confidence score, batch-processed and exported to CSV.

Built it half as an ML exercise, half because I was curious what my own resume looks like from the other side of the pipeline.

Don’t take my word for it — the receipts are on GitHub

02 · The Pack · Crew Duties

My Contribution

Cleaned a labeled resume dataset (sourced from HuggingFace’s resume-atlas) — stripping emails/phones/URLs, normalizing category names, and dropping short or duplicate entries.
Engineered handcrafted NLP features on top of TF-IDF/count vectorization — education, experience, and skill-keyword counts — to give the classifier more than raw word frequency to work with.
Trained and compared five classifiers — Naive Bayes, Logistic Regression, SVM, Random Forest, and KNN — and kept the best performer (Random Forest, ~82% accuracy in local testing) rather than assuming the fanciest model would win.
Built the hybrid scoring layer: a configurable α/β weighting between the classifier’s category confidence and the TF-IDF/cosine similarity score against the job description.
Built the Streamlit UI end to end — batch upload, adjustable pass/fail thresholds, per-resume drill-down, and Plotly visualizations — plus PDF/DOCX/TXT text extraction so it accepts resumes as people actually send them.

03 · In Bloom · Bright Stars

Key Features

40+ job categories

Resumes are automatically classified into one of 40+ categories rather than treated as one generic bucket.

Hybrid scoring

Classifier confidence and TF-IDF/cosine similarity to the job description are blended into one adjustable score, not just a single metric.

Batch processing

Screen a stack of resumes at once instead of one at a time, with exportable CSV results.

Format-agnostic input

Accepts PDF, DOCX, and TXT resumes directly — no manual copy-pasting into a text box.

04 · Sunrise · Planet-rise

Takeaways

The classifier was the easy 80%. The genuinely hard part was the hybrid scoring — a resume can be a perfect category match and a mediocre fit for one specific job description, or vice versa, and collapsing that into one number without losing the nuance took more tuning than the model training did. Also: building a resume screener while job-hunting is a slightly surreal form of dogfooding.

Previous waypointPrevious burn · 3 of 10

MLXnarrator

Give screen recordings a voice, offline on a Mac

Next waypointNext burn · 5 of 10

ArtAttack

A satirical first-person art-gallery vandalism sim

All Projects