title	Candidate Ranking Demo
emoji	🏆
colorFrom	blue
colorTo	green
sdk	docker
app_port	7860
pinned	false

Redrob Data & AI Challenge — Candidate Ranking System

A candidate ranking system for the Redrob Data & AI Challenge. Given a job description and a pool of candidate profiles, the system scores and ranks candidates by how well they fit the role.

Architecture

The system uses a two-phase architecture:

• Phase 1 — Offline pre-computation. Heavy, reusable work that does not depend on a specific query is computed once and cached to artifacts/. This is split into two tracks:
  • Track 1 — Structured signals. Rule-based scores derived from structured candidate fields: a hard filter, availability, and credibility.
  • Track 2 — Semantic representation. Builds a text representation of each candidate and encodes it into embeddings for semantic matching.
• Phase 2 — Ranking. Combines the pre-computed Track 1 signals and Track 2 embeddings against a given query to produce a final ranked list of candidates.

All field access goes through field_map.py, the single source of truth for candidate field paths. Scoring scripts never read candidate fields directly.

Interactive demo (Streamlit / Hugging Face Spaces)

A Streamlit app runs the full pipeline end-to-end on an uploaded candidate pool (or a built-in sample), with no pre-computed artifacts:

pip install -r requirements.txt
streamlit run app.py

Upload a JSON array or JSONL file of up to 100 candidate objects (same schema as data/sample_candidates.json), or click Use Sample Data. The app scores Track 1 + Track 2 in memory, ranks the pool, shows the scores as an interactive table, and offers a ranked-CSV download. The demo logic lives in app.py and demo_pipeline.py; both reuse the same scoring functions as the offline pipeline.

Setup

pip install -r requirements.txt

Place input data under data/. Pre-computed outputs are written to artifacts/.

How to run

Phase 1 — offline pre-computation

Run the Track 1 (structured) and Track 2 (semantic) steps to populate artifacts/:

# Track 1 — structured signals
python phase1/track1_hard_filter.py
python phase1/track1_availability.py
python phase1/track1_credibility.py

# Track 2 — semantic representation
python phase1/track2_text_builder.py
python phase1/track2_embedding.py

Phase 2 — ranking

python phase2/rank.py

Dependencies

• numpy
• pandas
• sentence-transformers
• torch
• scikit-learn
• pyyaml
• tqdm

See requirements.txt.