Gridlock Hackathon 2.0 · Round 2 · Theme: Event-Driven Congestion (Planned & Unplanned)
PulseGrid turns Bengaluru Traffic Police's raw incident stream into decisions. The moment an event is logged, it forecasts the disruption and recommends a deployment plan — how long it will take to clear, whether it needs barricading or a diversion, and how many officers to send — then plans city-wide staffing from a volume forecast and lets a controller run the whole thing live.
Built on 8,173 real, anonymised ASTraM incidents (Nov 2023 – Apr 2024). 🔴 Live demo → https://pulsegrid-diecjwaeyxc8supyngnd4y.streamlit.app/ · all ML in scikit-learn, no heavy dependencies.
- 🛰 A live command center, not a static dashboard. A discrete-event simulation replays Bengaluru in accelerated time — incidents arrive on the city's real per-hour rates, get scored by the models on the spot, and consume a finite officer pool. Watch the City Pressure Index move, the force saturate, and incidents queue as awaiting units.
- 💬 You can just ask it. A natural-language co-pilot answers "how many officers for a festival on Hosur Road Saturday evening?" using the same models — fully offline, no API key.
- 🧠 It proves its own worth. A counterfactual engine shows PulseGrid's dispatch logic gets units to critical incidents ~50% faster than a reactive control room, on identical officers.
| Pain point (from the theme) | What PulseGrid does |
|---|---|
| Event impact is not quantified in advance | Clearance-time + priority + closure models predict impact at intake |
| Resource deployment is experience-driven | Closure model + a load-balancing optimiser produce a data-driven plan |
| Planned events aren't pre-planned for | The Plan Event view scales officers/barricades/advisory window by crowd size |
| Triage under a surge is first-come | Severity-aware dispatch + a forecast-aware reserve (proven in Policy Lab) |
| No post-event learning loop | Every model trains on resolved events — the resolution log is the label |
| Model | Task | Result |
|---|---|---|
| Road closure | Will this event need a barricade / diversion? | AUC 0.787, PR-AUC 0.42, Brier 0.08 (well-calibrated) |
| Priority | High vs Low triage | AUC 0.9997 — recovers ASTraM's rule instantly |
| Clearance time | How long until cleared? | P50 median err ≈35 min + conformally-calibrated 90% band (verified coverage) |
| Daily volume | City-wide event load | MAE 18/day, beats the naïve baseline (21) |
Proactive advantage: at its alert threshold the closure model surfaces ~57% of all road-closure events while flagging only ~16% of incidents — a 3.6× lift over reviewing cases at random.
Every claim is validated: the clearance band is conformally calibrated to true 90% coverage, closure
probabilities are well-calibrated (Brier 0.08) with a published operating-point table, the volume
forecast beats a seasonal-naive baseline by ~15%, and the Live-Ops simulator reproduces the real
hour-of-day incidence at r≈0.98. All ML is scikit-learn HistGradientBoosting — no heavy deps.
- 🛰 Live Ops — real-time engine: incidents stream in on real arrival patterns, are scored live, and draw from a finite officer pool. City Pressure gauge, live map, force-utilisation, congestion-ripple board, dispatch log, and a 💥 Inject button for live what-if drama.
- 💬 Ask the Console — plain-English co-pilot over the whole system (event staffing, corridor risk, hotspots, forecast). Offline intent parser, no external API.
- ⚡ Live Triage — log an event → a colour-coded dispatch ticket (clearance band, priority, closure probability, officer count, barricade + diversion) with a "Why this call" explainer.
- 📋 Plan Event — pre-deployment plan for planned events (festival, rally, VIP, procession, roadwork), scaled by expected crowd, with the venue's historical busy-hour profile.
- 🗺 Hotspot Map — every incident on a dark Bengaluru map, severity-coloured and filterable; surfaces the real chokepoints (Silk Board, Mekhri Circle, Yeshwanthpura…).
- 📈 Forecast & Staffing — 7-day volume forecast with surge alerts + a per-corridor officer allocation you can size to any headcount.
- 🧠 Policy Lab — replays one event-surge stream under a reactive control room vs PulseGrid's severity-aware dispatch: ~50% faster to critical incidents (≈51 → 24 min on a 16-officer force), with sliders to vary force size and surge intensity.
- 🧪 Model Card — validated metrics, methodology, and an honest limitations note.
pip install -r requirements.txt
streamlit run app.pyThe dataset lives at data/astram_events.csv (the loader also finds it at the repo root or under any
ASTraM-style filename). Models train once on first load (~1 min) and are cached.
pulsegrid/
├── app.py # Streamlit command console — 8 tabs (UI)
├── core.py # data prep, 4 models, recommender, planner, forecast, optimiser
├── sim.py # real-time simulation engine + Policy Lab counterfactual
├── assistant.py # offline natural-language co-pilot
├── data/astram_events.csv
├── requirements.txt
├── .streamlit/config.toml
├── APPROACH.md # methodology & design write-up
└── CHANGELOG.md # what shipped, and what's new
Clearance time has a heavy tail (a few events span days), so we report median error and present a clearance band rather than overclaiming a point estimate. Priority is near-deterministic from cause in this dataset — the model recovers that logic at ~100%, which is a feature (consistent instant triage), not leakage. The Policy Lab compares two dispatch policies on an identical simulated stream; it is a controlled what-if, clearly labelled as such, not a claim about historical outcomes.
@gitadi2 ·
@adisatapathy ·
satgriezeleo1007@gmail.com
Built for Flipkart Gridlock 2.0, 2026
