SkillFit Analytics

A system for measuring skill transferability between jobs using deep metric learning on 600,000+ Indian job postings.

Inspired by:

• Technology and the Shifting Architecture of Occupational Skills
• From Posting to Prediction: Building Validated Workforce Analytics

---

What it does

Given any job description, SkillFit embeds it into a 128-dimensional skill space and finds the most similar roles in the dataset — quantifying how transferable skills are between occupations.

Two tools:

• Skill Match Comparator — paste two job descriptions, get a 0–100 compatibility score
• Occupation Predictor — paste a job description, get its predicted occupation group and nearest matches

---

Architecture (Merge of 2 papers Architecture)

Raw job text
    │
    ▼
Jina Embeddings v3 (ONNX)  →  1024-dim semantic vector
    │
    ▼
AttnNet  →  aggregates K postings into one prototype vector
    │
    ▼
CompressNet  →  projects 1024-dim → 128-dim skill vector (L2 normalized)
    │
    ▼
Cosine similarity against 600k trained vectors  →  occupation prediction

Training: Stage 1 AttnNet warmup (MSE vs centroid, 6 epochs) → Stage 2 end-to-end (35 epochs, loss = 0.3×triplet + 0.7×SupCon + 0.01×center, cosine margin=0.8)

Dataset: 600,561 job postings across 180 occupation groups

---

Results

Metric	Baseline (Jina)	SkillFit	Δ
Recall@1	0.6279	0.6631	+0.0352
MRR	0.7169	0.7453	+0.0284
Silhouette	−0.0330	0.0405	+0.0735
Separation	−0.1359	0.5860	+0.7219

125 of 180 occupation groups improved in cluster silhouette score.

---

Stack

Layer	Technology
Embeddings	Jina Embeddings v3 (ONNX Runtime)
Model	PyTorch (AttnNet + CompressNet)
Nearest neighbor	torch.topk cosine similarity
Backend	FastAPI + Uvicorn
Frontend	Vanilla JS / HTML / CSS
Deployment	Google Cloud Run (backend) · Vercel (frontend)
Artifact storage	Google Cloud Storage

---

Running locally

Backend

cd backend
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt
python server.py

Requires outputs/ and jina-embeddings-v3/ to be present locally.

Frontend

# Any static server, e.g.:
python -m http.server 3000

Open http://localhost:3000. The frontend reads config.js for the backend URL — leave it as-is for local dev (falls back to http://localhost:8000).

---

Deployment

• Backend → Google Cloud Run (Dockerfile in backend/, artifacts pulled from GCS at build time)
• Frontend → Vercel (reads vercel.json, backend/ is ignored)

After deploying, set your Cloud Run URL in config.js:

window.BACKEND_URL = "https://your-service.run.app";