SAARTHI — Bengaluru Traffic Ops Console

SAARTHI

Astram · Event-Driven Congestion Console

—

Incidents analysed

—

Congestion-min mapped

—

Risk capture @top20%

—

Congestion preventable

Deployment plan · by day

—

of expected congestion-minutes
covered by 8 patrol anchors

What-if — re-optimise live

Units (K)8

📍 Drop custom event

Marginal coverage — how many units?

– → 50%– → 70%– → 80%

Risk heat

Units

Hotspots

Barricades

Spillover

Blind spots

low

high

Pick an event above to re-optimise the deployment live. Or open Spillover and click a corridor for its cascade + diversion.

Would it have helped?counterfactual

We don't just predict — we back-test the plan against history. If SAARTHI's units had been pre-positioned, how much congestion was preventable?

Reach equity — central zones hog coverage; periphery is starved. We surface it, not hide it.

Barricading planPS ask ✓

Where to physically place barricades for closures — real approaches + junctions, with the diversion to turn traffic onto. Dot = response urgency.

Spillover & diversionPS ask ✓

Mined cascades within 90 min. Click a row to trace it on the map.

Click a corridor to see total system impact (direct + cascaded).

Emerging-event alarm

Spatio-temporal bursts (≥4 incidents, 1.5 km, 90 min) — sudden gatherings & weather. Caught the Mar-7 city-wide rain flooding live.

Post-event learningPS gap ✓

Walk-forward: re-fits each week on past data only, scored on the next unseen week. Capture holds & rises — the nightly loop, proven.

wk1 · –mean –wk– · –

Clearance hold-timemanpower ✓

P50 typical · P90 plan-for. Dot = response urgency tier.

≤15 min ≤1 hr standard

Planned-event playbook

Archetypes parsed from field reports · median congestion-min

⚑ Two traps we defused

Time is corrupted

Raw hour-of-day peaks at 2 AM — a timezone + batch artifact. We drop it; key on day-of-week. (Same instinct keeps batch entries out of the alarm above.)

Clearance time is mostly fake

Batch auto-closes pile into a synthetic 2–3 h band. We flag them as censored and learn only from real signal.

SAARTHI · validated on 30 unseen days · learning proven over 13 weeks