v2.6 now available

For Northeastern co-op students

Find the co-ops that actually fit your resume.

NUWorks Grader reads every posting the way you would — then scores it against your skills, flags the ones you're not eligible for, and saves the strong matches. Optional Semantic AI matches by meaning, too. All in your browser, all private.

Add to ChromeIt's free Try the interactive demo

Explore fictional roles with the real lexical matcher. No account or installation.

100% Local & private

250+ Skills recognized

On-device Semantic AI (opt-in)

Free And open source

See your next co-op more clearly.Play the walkthrough

Where our users land offers

A clearer view of every opportunity

Match scores and eligibility badges, right where you browse. Spot the roles that fit before opening another posting.

NUWorks Extension Interface Discovery Page with Color-Coded Badges

NUWorks Extension Interface Search Page with Color-Coded Badges

Select any screenshot for a closer look.

A redesigned Job Explorer and popup

Match scores, an explainable breakdown for every job (now including a Semantic AI bar), smart filters, and a calm interface that follows your system light or dark theme.

New

A job's match breakdown with sub-scores and matched skills

Why use this extension?

Stop wasting time on jobs that don't match your profile. Let smart, on-device analysis do the heavy lifting.

Auto-grading & visual badges

Instantly see match scores, external-application warnings, and eligibility status with color-coded badges on the job list.

Smart filters & batch actions

Filter by match score or freshness. Save all matched jobs in one click, or clean up your saved list just as fast.

Resume parsing

Upload your PDF resume. We extract the text and skills locally to find your best matches.

Semantic AI matching Opt-in

A 33M-parameter BGE model fine-tuned for resume–job fit runs on your device and scores each posting by meaning. Match becomes 65% keyword + 35% AI, with the split shown on every card.

Implied skills

Listing React implies JavaScript; PyTorch implies Python. Related skills earn partial credit and show up as dotted chips, so you're not penalized for not spelling everything out.

Faster Job Explorer

Scores render from a single listing call — no per-job fetching. Extra details load in the background with a status chip, and the AI refinement never blocks the page.

On-device matching · Opt-in

Semantic AI trained for resume-to-job fit

Turn on Semantic AI and a compact model fine-tuned on labeled resume–job pairs runs entirely on your device, re-reading every posting to score how well it fits by meaning—not just keyword overlap.

The Job Explorer with Semantic AI enabled, showing a blended match score plus a Keyword · AI split on every card

Model release · October 15, 2026

Broader coverage, stronger test rankings

Development and evaluation used 864 original fictional resumes spanning 24 majors and six career stages, including 264 profiles with skills outside their named major. OpenAI teacher models scored the resume–job matches.

Updated model. The release is available on Hugging Face. Existing users can update their downloaded model in Semantic AI.

Mean per-resume Spearman 0.7682 published baseline: 0.4221 +0.3461 rank correlation

NDCG@10 0.9408 published baseline: 0.8644 Better top-10 ranking agreement

Held-out benchmark: 2,304 pairs from 144 fictional resume documents, 24 scenario families and 24 unseen job scenarios. Training, validation and test sets kept resume families and job scenarios separate. The 95% family-bootstrap interval for the ranking gain is +0.2832 to +0.4136, conditional on this job pool.

Real-resume check: across 960 pairs from 24 resumes and 40 jobs, ranking correlation improved from 0.6812 to 0.7978 under the updated grading rubric. The 95% interval for the gain was +0.0937 to +0.1428. The original rubric also showed a ranking gain, from 0.6838 to 0.7907.

Numeric grades remain mixed. Mean error on this real-resume check rose from 13.06 to 13.36 points under the updated rubric and from 12.50 to 13.92 under the original rubric. On earlier tests, it fell from 36.01 to 17.69 on the fictional benchmark and from 11.93 to 11.16 on two legacy resumes. Each resume contributes equally to these averages.

Combined score check: with the default keyword and AI weighting, mean grade error fell from 15.38 to 14.01 points on the updated real-resume rubric in a Node simulation. Ranking results were uncertain: the top-10 metric declined slightly, and both ranking-change intervals included zero.

Limits of the evidence: these scores measure agreement with AI grading rubrics, not hiring outcomes or proof of improvement for every major. The fictional benchmark has only one held-out scenario family per major. The current semantic grade scale tops out at 70; this is separate from the extension’s combined keyword and AI score. Browser execution remains unverified. Read the full evaluation · Download the aggregate results

Keyword layer

The classic scorer checks required skills, keywords, and eligibility — fast, transparent, and always on. It now also credits implied skills: React on your resume counts toward a JavaScript requirement.

Semantic layer

mdbr-leaf-mt-resume-grader is the released 33.36M-parameter BGE-small fine-tune used by the extension. It embeds your resume and each posting, then measures how close they sit in meaning. “Built REST services in Flask” counts toward “backend web development”.

Blended score

Your final match is 65% keyword + 35% semantic. Every card shows the split (Keyword 96 · AI 79) and the details view adds a dedicated Semantic (AI) bar so you can see exactly what moved the number.

Explainable, down to the bar

Open any posting and the match breakdown shows Skills, Keywords, and Semantic (AI) side by side. Hover a card's split to see the raw cosine similarity behind the AI number.

Private by design. Your resume and every job description are embedded in a Web Worker in your browser. Nothing is uploaded.
Lightweight until you opt in. The extension ships without the model. Ticking “Semantic AI” downloads the runtime and model once (~57 MB) and keeps them cached.
Scores never wait on it. Keyword scores appear instantly; the AI refinement streams in behind a “Refining scores” chip and never blocks the page.
Honest about hardware. Before enabling, you're told it's compute-intensive on MacBook Airs, Chromebooks and older laptops. Thread count auto-scales to your CPU.