Flipkart Gridlock 2.0 · Round 2 · Prototype

PWCA

Project Without a Cool Acronym

We don't just detect congestion —
we price it, then relieve it.

1 Find→ 2 Price→ 3 Predict→ 4 Dispatch→ 5 Prove

Decision-intelligence for parking-induced congestion · Bengaluru Traffic Police

→ next · F fullscreen · ? help

!

The Problem01 · THE GAP

A ₹500 fine hides a
city-scale cost.

Illegal parking doesn't just block a kerb. It steals road capacity → queues form → drivers idle → fuel burns → the air fouls. BTP logs every violation — but the log never adds up the damage.

The visibility gap

BTP has the data. What's missing: where it hurts most, what it actually costs, and where to send the next scarce officer.

382

City PPI

Parking-Pressure Index · Severe

₹95,825

Cost / day

≈ ₹3.50 cr a year

17,179

Veh-min lost / day

people stuck in the queue

398 kg

CO₂ / day

+ a PM2.5 health bill

1,698

Hotspots

across 54 police stations

298,445

Violations

Nov 2023 – Apr 2024

Source — BTP parking-violation log (anonymised), 151 days. Modelled estimates; every figure traces to a formula + cited parameter.

The Fork02 · APPROACHES

Three ways to fight it. Only one scales.

Approach A · Detect more

More cameras

Bolt ANPR / computer-vision cameras onto junctions to catch more violations.

✗ No imagery in this data — and detection ≠ relief.

Approach B · More boots

Blanket enforcement

Flood every market with officers and tow-trucks until it clears.

✗ Officer-hours are scarce; untargeted spend burns them.

Approach C · Our approach

Price it, then optimise

Convert every violation into ₹ of delay, then spend each scarce officer-hour where it buys the most relief.

✓ Targets the 5% of spots that cause 76% of the cost.

North-star

Vehicle-minutes of congestion relieved per officer-hour. Everything is measured against it.

The Platform03 · SOLUTION

One pipeline — from raw log to proven relief.

1

Find

298K violations → 1,698 hotspots, ranked by ₹ not counts.

2

Price

Attribution engine: violation → delay → ₹, fuel, CO₂, health.

3

Predict

Backtested forecast of tomorrow's pressure — beats baselines.

4

Dispatch

ROI optimiser + patrol route for scarce officer-hours.

5

Prove

Every clear booked to an auditable Impact Ledger.

PWCA command overview — **Fig. 01** — Command overview: the five-step story, in one screen.

1

Step 01 · Find04 · THE PIPELINE

Where the congestion
really is.

Every violation snaps onto an Uber-H3 grid and collapses into 1,698 ranked hotspots — sorted by the rupees of delay they cause, not by raw ticket count.

1,698

Hotspots

76%

of cost from top 5%

13

spots = half the cost

Hotspot map of Bengaluru — **Fig. 02** — Hotspots ranked by ₹/day, mapped over the city.

2

Step 02 · Price05 · THE PIPELINE

Turn each violation
into rupees.

The Congestion Attribution Engine runs the causal chain — and every number traces to a formula + a cited parameter in params.yaml.

Violation→ Capacity lost→ Vehicles delayed→ ₹ time · fuel · CO₂ · health

₹95.8k

priced / day

398 kg

CO₂ / day

₹2.38 L

health / yr

City insights and cost concentration — **Fig. 03** — The cost is concentrated: fix 13 spots, halve the problem.

3

Step 03 · Predict06 · THE PIPELINE

Tomorrow's pressure,
today.

A gradient-boosted, time-series-backtested model ranks where pressure flares next. It's only credible because it beats the naïve baselines — and it does, with no leakage.

0.477

Precision @20

vs 0.385 / 0.335 naïve

+24%

lift over baseline

31

backtest days

Next-day forecast with baseline comparison — **Fig. 04** — Next-day forecast, scored against named baselines.

4

Step 04 · Dispatch07 · THE PIPELINE

Spend scarce officer-hours
where they pay.

An ROI optimiser — cost × risk, capacity-constrained — writes the work orders and the patrol route, then proves the lift over simply deploying to the busiest spots.

₹36.7k

relief / day

≈ ₹1.34 cr / yr

+7.2%

vs busiest-spots

~34×

vs random deploy

Enforcement optimizer and work orders — **Fig. 05** — Work orders, ranked by relief — dispatched to the field.

5

Step 05 · Prove08 · THE PIPELINE

Every action,
measured & auditable.

Each clear is booked to an append-only Impact Ledger — rupees, vehicle-minutes and CO₂, traceable to the formula and the hotspot it came from. No vanity metrics.

Live

SQLite + SSE ledger

₹ · min · kg

booked per action

0-install

WhatsApp dispatch

Impact ledger with booked relief actions — **Fig. 06** — The Impact Ledger: relief booked, with provenance.

The Differentiator09 · TWIN RACE

Same officers. Same hours.
4.7× the relief.

One simulated day · 12 officers · 180 officer-hours. Random chases active hotspots; PWCA follows the optimiser. The relief diverges.

PWCA optimiser₹18,862

Random / by-count₹3,983

₹105

per officer-hr · PWCA

₹22

per officer-hr · random

The north-star, made visible.

Twin Race counterfactual result — **Fig. 07** — Two lanes, one day: cumulative cost (lower is better) splits apart.

What Changes10 · IMPACT

From reactive & blind to targeted, measured — and self-funding.

Reactive patrols → ROI-ranked work orders for the next shift.
A ₹500 fine → an itemised, auditable congestion bill.
Invisible cost → a live ledger of relief actually delivered.
Cost centre → recoverable municipal revenue.

Polluter-Pays · Elite Junction

Its parking spillover costs the city ₹65.29 L / year. 30% is recoverable as a congestion levy — ₹19.59 L / year — alongside a Gemini-drafted official notice (with deterministic fallback).

Polluter-Pays itemised congestion bill — **Fig. 08** — An itemised ₹ bill: time, fuel, health, carbon → recoverable revenue.

Make it Tangible11 · SIMULATOR

Watch one parked car
choke an arterial.

A kinematic-wave digital twin: set how many vehicles, of what type, under what demand, and the queue backs up the road in real time at a modelled shockwave speed — magnitudes anchored to the Attribution Engine. Then drag TOW to clear the lane and book the relief.

₹17.3k

priced / day here

52%

capacity lost

16 km/h

shockwave upstream

Butterfly-Effect kinematic-wave simulator — **Fig. 09** — The jam backs up in accelerated time; clearing relieves **₹15.9k/day**.

Ease of Use12 · ASK & ACT

Type plain English.
Get a priced action — you confirm.

The docked copilot plans & drafts; you approve. Gemini parses intent + slots only — the server runs the trusted engines and owns every number. Nothing changes until you confirm.

Plain-English ask→ Gemini · intent + slots→ Trusted engine→ Priced card · you confirm

Anti-hallucination · works with no API key

“Send 6 officers to the top hotspots” → a dispatch plan relieving ₹34.6k/day, +7.8% over chasing the busiest spots. Every figure from the optimiser; a deterministic parser is the floor when there's no key.

Ask & Act agentic copilot proposing a dispatch plan — **Fig. 10** — One box → a typed, confirmable card: plan · route · notice · navigate · answer.

Under the Hood13 · CREDIBILITY

Built to be audited, not just demoed.

The stack

Python · pandas Uber H3 (res 9) XGBoost OR-Tools / PuLP FastAPI SQLite · PostGIS React · Vite Leaflet · Recharts Gemini + fallback OSM → MapMyIndia

Pure-function pipeline → artifacts → API → views. The dataset (≈110 MB, 298K rows) processes end-to-end in ~18 seconds.

Why you can trust the numbers

Every figure = formula + cited parameter in one auditable surface.
Low / expected / high bands; modelled values labelled estimates.
Time-series splits, named baselines, a model card per model — no leakage.
Runs with no API key (deterministic fallback); tri-lingual EN · हिंदी · ಕನ್ನಡ.

What's Next14 · ROADMAP

From prototype to a product in officers' hands.

Now · Prototype

Decision platform

Find→Prove on historic BTP data
Twin Race counterfactual
Polluter-Pays bills

Next · Field app

Officer mobile app

Live CRITICAL alerts
One-tap dispatch
"CLEARED" confirms → ledger

Then · Live data

Real-time feeds

ANPR / sensor ingest
MapMyIndia road network
Continuous re-optimise

Scale · Policy

City & citizens

BBMP / DULT dashboard
Citizen cost + accountability
Next cities

The app ships the same engine — only the surface changes: from a war-room screen to a constable's phone.

PWCA · Bengaluru

PWCA

Price the congestion. Relieve it. Prove it.

Find→ Price→ Predict→ Dispatch→ Prove

Flipkart Gridlock 2.0 · Round 2 · Prototype Phase

We don't just detect congestion —we price it, then relieve it.

A ₹500 fine hides acity-scale cost.

Three ways to fight it. Only one scales.

More cameras

Blanket enforcement

Price it, then optimise

One pipeline — from raw log to proven relief.

Find

Price

Predict

Dispatch

Prove

Where the congestionreally is.

Turn each violationinto rupees.

Tomorrow's pressure,today.

Spend scarce officer-hourswhere they pay.

Every action,measured & auditable.

Same officers. Same hours.4.7× the relief.

From reactive & blind to targeted, measured — and self-funding.

Watch one parked carchoke an arterial.

Type plain English.Get a priced action — you confirm.

Built to be audited, not just demoed.

From prototype to a product in officers' hands.

Decision platform

Officer mobile app

Real-time feeds

City & citizens

Price the congestion. Relieve it. Prove it.

We don't just detect congestion —
we price it, then relieve it.

A ₹500 fine hides a
city-scale cost.

Where the congestion
really is.

Turn each violation
into rupees.

Tomorrow's pressure,
today.

Spend scarce officer-hours
where they pay.

Every action,
measured & auditable.

Same officers. Same hours.
4.7× the relief.

Watch one parked car
choke an arterial.

Type plain English.
Get a priced action — you confirm.