role-radar-scoring

Role Radar — Job-Profile Fit Scoring Models

Local-first two-stage scoring pipeline for job-candidate fit prediction. Trained on Indian job market data across 15 role families.

GitHub: somus/role-radar | Dataset: oksomu/role-radar-dataset

Models

Architecture

LightGBM scores all jobs (<1ms each)
  → Composite = skills×0.5 + location×0.25 + seniority×0.15 + domain×0.10
  → Shortlist: composite > 60
    → MiniLM reranker reranks shortlist (~14ms/pair)
      → Final ranked feed

Metrics (108 gold eval pairs)

Latency (M1 Pro)

Training Data

• 2,500 jobs (LinkedIn India, 12 cities, 15 role families)
• 640 profiles (438 synthetic + 202 real)
• 22,465 labeled pairs (Gemini 2.5 Flash Lite teacher)

Usage

All models use ONNX Runtime:

import onnxruntime as ort

# LightGBM
session = ort.InferenceSession("skills_lgbm.onnx")
score = session.run(None, {"features": feature_vector})[0]

# Reranker
session = ort.InferenceSession("reranker.onnx")
logits = session.run(None, {"input_ids": ids, "attention_mask": mask})[0]

Files

skills_lgbm.onnx              — LightGBM skills scorer
skills_calibration.json       — isotonic regression (raw → 50-100 scale)
embedding_encoder.onnx        — MiniLM encoder for cosine sim feature
embedding_encoder.onnx.data   — encoder weights
embedding_tokenizer/          — tokenizer for encoder
reranker.onnx                 — MiniLM cross-encoder
reranker.onnx.data            — reranker weights
reranker_tokenizer/           — tokenizer for reranker
config.json                   — feature names, composite weights, thresholds

Score Bands