parallax

About

European municipalities have access to billions in EU funding but lack the resources to navigate hundreds of grant programmes. Parallax matches a city's development plan against available grants using semantic search, then explains each match through SHAP/LIME, counterfactual analysis, and LLM-generated narratives: tailored to three audience levels (citizen, analyst, auditor).

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Usage

Build embeddings and run matching:

python -m src.preprocessing.pipeline

Build knowledge graph:

python -m src.knowledge_graph.builder

Launch the app:

uvicorn src.app.api:app --reload

Project structure

src/
├── models.py             # data models
├── data_collection/      # data loading
├── preprocessing/        # text cleaning, chunking
├── retrieval/            # embeddings, semantic matching
├── knowledge_graph/      # domain graph
├── xai/                  # SHAP, LIME, counterfactual, RAG
├── analysis/             # gap analysis, bias detection, confidence
└── app/                  # FastAPI + frontend

Company: EY — AI Explainability Challenge