---
title: Redrob Intelligent Candidate Ranker
emoji: 🧭
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: mit
---

# Redrob Intelligent Candidate Ranker - live demo

Evidence-based AI candidate ranker for the India Runs (Redrob) Track-1 challenge.
This Space runs the **full hybrid pipeline live** - BM25 + dense `bge-small` embeddings
fused with RRF, then a 7-component signal scorer + trap gate, with grounded reasoning.

Paste/edit the target JD, hit **Rank candidates**, and inspect the per-candidate score
breakdown. Toggle **Inject a keyword-stuffer** to watch the trap gate demote it.

Code + methodology: https://github.com/DhruvGoyal404/redrob-ranker

## Architecture & design

<!-- HF rejects binary files in git, so these render from the GitHub copy (inline, not links). -->

**System architecture** - offline embedding precompute feeding the CPU-only online ranking pipeline:

![System architecture diagram](https://raw.githubusercontent.com/DhruvGoyal404/redrob-ranker/main/diagrams/architecturediagram.png)

**Module structure** (class diagram):

![Module class diagram](https://raw.githubusercontent.com/DhruvGoyal404/redrob-ranker/main/diagrams/classdiagrams.png)

**Use cases**:

![Use case diagram](https://raw.githubusercontent.com/DhruvGoyal404/redrob-ranker/main/diagrams/usecasediagram.png)

> This file is the Hugging Face Space config (the frontmatter above). It is intentionally
> separate from the repo's main README.md so the GitHub README stays clean.