01
Prediction creates the possibility.
LSTM reads the last hour as a sentence. It remembers the school-run spike, the secretariat pulse, the rain that turns MG Road into a parking lot. Without memory, routing is gossip.
Compiling the hour
Compiling the hour
The premier intelligence layer for urban movement
That’s traffic.
Manifesto
Every commuter served.
Every corridor priced in time.
Every signal timed.
Every vehicle counted.
Training creates the possibility. Inference — here, the live prediction of a city’s next hour — creates the value. The more we ask of mobility, the more prediction it takes. And prediction is becoming the bottleneck. Traffic was built for this.
Live graph runtime
We pioneered a hybrid stack: temporal memory, spatial attention, camera density, and a routing policy that updates as the city breathes. Ember marks the vehicles. Graphite holds the roads. The rest is silence.
Loading graph runtime
15.4k
Intersections
35.3k
Road segments
4.5 km
OSM radius
0.9766
LSTM R²
10–25%
Travel-time cut
TVM
Kerala capital
Thesis · proof · evidence
01
LSTM reads the last hour as a sentence. It remembers the school-run spike, the secretariat pulse, the rain that turns MG Road into a parking lot. Without memory, routing is gossip.
02
Congestion does not live on a single edge. It hops. GAT attention is how a jam at East Fort writes itself onto Kowdiar twelve minutes later — in public, as weights.
03
PPO does not compute the shortest path. It learns the cheapest future. Dijkstra is a photograph. The agent is a rehearsal.
What you can do here
Choose date, time, and scenario (rain, accident, normal). See predicted speeds at named places like Palayam, Technopark, and Kowdiar. Then ask PPO for a start–end corridor.
Open Predict →
Color the full Thiruvananthapuram road graph by predicted speed. Pick two places, generate a route, and compare shortest vs fastest vs balanced paths.
Open Map →
Focus on one place. See which nearby roads are strongly linked in the model — a heat map of connections, not jam colors.
Open Influence →
Upload a road photo. YOLOv8 counts cars, buses, bikes, and trucks so density can inform the next forecast.
Open Detect →
About the speeds
Speeds on Predict, Map, and Influence come from a deterministic demo predictor: same date, time, and scenario always produce the same road speeds. Rain and rush hour make corridors slower on purpose. The road layout and place names come from real OpenStreetMap data for Thiruvananthapuram. YOLO detection on uploaded photos is live for that image. Together, the site is a working lab for the capital’s graph — not a live CCTV feed of every junction.
Case
The graph is Palayam, MG Road, Kowdiar, Pattom, and the coastal approach — extracted with OSMnx, simulated with weather, density, and the ugly truth of peak hour. Against Dijkstra, the PPO agent cut travel time 10–25%. Against a lone LSTM, the hybrid model traded a little MAE for a map that actually routes.
“
The first time a dashboard told us a jam would exist before the cameras filled, we stopped arguing about sensors and started arguing about policy.
“
It looks like a lab notebook because that is what a city is — a long experiment that should not surprise the people who live in it.
A morning in the capital
Traffic is built around Thiruvananthapuram’s real geography: Fort markets, Palayam offices, Kowdiar residences, the medical college belt, Kazhakkoottam’s IT corridor, and the coastal approach to the airport.
MG Road and Palayam tighten. Predict with morning time + normal or heavy scenario to see place speeds drop.
Switch scenario to rain, map Airport → Secretariat, compare shortest vs fastest. Time and distance diverge.
Technopark → Kesavadasapuram under heavy traffic. Influence near Kazhakkoottam shows which feeder roads are coupled.
Upload a junction photo on Detect. Counts are live for that image; city speeds stay demo forecasts.
Clear scenario after 22:00 raises predicted speeds. Useful to show the model responds to clock and weather knobs.
Do not present demo speeds as live Google traffic. Say: real OSM graph + repeatable scenario forecasts.
Common questions
Road layout and places are from OpenStreetMap. Speeds on Predict/Map/Influence are demo forecasts from date, time, and scenario. YOLO on Detect is live for the photo you upload.
Predict a scenario, open Map with two places, then Influence near one of them. That three-step loop shows forecast → route → road links.
It is a real capital graph with secretariat pulse, coastal approaches, hills, and an IT corridor — harder and more honest than a toy grid.
LSTM R² and PPO travel-time cuts are from the research track. The interactive site demonstrates the same loop with a lighter demo predictor so anyone can run it.
Start
Open the studio. Pick a corridor. Watch the graph attend. Then give the agent a destination.