Gridlock Hackathon 2.0 — Theme 2 (Event-Driven Congestion) · Target: Bengaluru Traffic Police (BTP) & Flipkart

This system forecasts the congestion impact of event-driven disruptions and recommends proportional officer deployment. Built on 8,173 real-world BTP incident records — no synthetic data.

See METHODOLOGY.md and FEATURE_AVAILABILITY.md for the full data audit, scoring logic, and statistical disclosures.

🌐 Live App · 📹 Demo Video · 💻 GitHub

---

What It Does

Feature	Description
Live Disruption Map	Stratified sample of historical incidents across Bengaluru, color-coded by severity
Response Recommendation Engine	Rule-based prediction using only report-time data (Priority + Corridor Frequency) — no post-resolution leakage
Post-Event Learning Replay	Chronological calibration replay showing predicted vs. resolved severity gap over time
Explainability Panel	Per-prediction breakdown of score components; flags corridor fallback inline if lookup misses
Data Limitations Drawer	Full-screen disclosure of model constraints, logging gaps, and ML confidence

Three-Signal Architecture

The system produces three independent severity reads per incident, never blended:

1. Historical Baseline (display_severity_bucket) — retrospective, based on actual median clearance time for the cause+corridor combination.
2. Report-Time Predictive Score — deterministic rule engine (Priority 57.14% + Corridor Frequency 42.86%), computable at the moment of logging.
3. ML Directional Read (ml_severity_bucket) — LightGBM classifier on 6 report-time features; labeled low-confidence and secondary.

Known Limitations

• Coarse Bucket Resolution: 4-bucket output (Low/Medium/High/Critical) built on 2 input signals only, to strictly avoid label leakage.
• Low ML Confidence: Macro-F1 of 0.477 on held-out test split. Kept explicitly secondary to the rule engine.
• Chronic Logging Gaps: ~61% of records lack completion timestamps — only ~39% of events have a resolvable duration for ground-truth labeling.

See METHODOLOGY.md for full statistical detail.

Screenshots

Explainability Panel & Fallback Disclosure Inline warning when corridor lookup misses — prevents silent zero	Three-Signal Architecture Panel Historical Baseline · Rule-Based Predictive · ML Directional Read
Post-Event Learning Replay Predicted vs. resolved severity gap plotted chronologically	Data Limitations & Disclosures Full-screen disclosure of model constraints and logging gaps

Dataset: data/astram_event_data.csv — 8,173 real-world BTP incident records. No synthetic or simulated rows in the training or calibration pipeline.

Data Pipeline & Artifacts

To run the pipeline from scratch:

cd project/data-pipeline
uv run python clean.py
uv run python severity_score.py
uv run python predictive_score.py
uv run python calibration_ledger.py
uv run python train_model.py
uv run python validate.py

Key artifacts generated:

• cleaned_data.parquet — standardized historical ledger
• predictive_cutoffs.json — corridor frequencies and rule-engine thresholds
• model.pkl — trained LightGBM model
• calibration_ledger.json & calibration_summary.json — Timeline Replay data

Setup

Backend

cd project
uv sync          # install all Python dependencies
uv run uvicorn backend.main:app --port 8000

Frontend

cd project/frontend
pnpm install
pnpm run dev

Environment Variables

The Next.js frontend proxies /api to http://127.0.0.1:8000 by default. To point at a remote backend:

# .env.local
BACKEND_URL=https://your-backend-url.com

Deploying the Backend (Google Cloud Run)

The backend is fully containerized.

# 1. Authenticate and set your project
gcloud auth login
gcloud config set project YOUR_PROJECT_ID

# 2. Build and deploy in one step (Cloud Build handles the Docker build)
gcloud run deploy gridlock-backend \
  --source . \
  --region asia-south1 \
  --platform managed \
  --allow-unauthenticated \
  --port 8080 \
  --memory 1Gi \
  --cpu 1

# 3. Copy the deployed URL and set it in frontend/.env.local
echo "BACKEND_URL=https://<your-service-url>.run.app" > frontend/.env.local

Note

The image is ~500MB — lightgbm and pandas are the heavy dependencies. --memory 1Gi is the minimum safe value for loading the model and parquet artifacts at startup.