Compiling the hour
Product
Traffic is not a dashboard bolted onto a model. It is a closed loop: see the road, remember the hour, attend to the neighborhood, then move the vehicle. Each surface is a chapter in the same lab book.
01
Forecast link speeds on any corridor in the Thiruvananthapuram graph. Switch LSTM and hybrid LSTM–GAT. Stretch the horizon from twenty minutes to two hours. The chart is not decoration — it is the contract the router will honor.
02
YOLOv8 reads CCTV the way a loop detector never could. Cars, buses, trucks, motorcycles. Counts become a density channel. When the camera disagrees with history, the model listens to the camera.
03
GAT does not average neighbors. It votes. The influence page is the vote made visible: which upstream edges are writing the future of the edge you care about.
04
PPO consumes predicted speeds as living edge weights. The policy is trained to minimize time, not distance. The recommended path can look wrong on a paper map and right on a clock.
What it is not
Consumer navigation is a reaction. It notices the jam you are already sitting in and offers you a slightly different jam. Traffic is a forecast with consequences. The model writes tomorrow’s speeds onto today’s graph. The agent is graded on whether people arrive.
The product surface is intentionally quiet. Parchment for thesis. Graphite for proof. Ember only when something must move. We borrowed that discipline from the laboratories that ship silicon, not from the marketing sites that ship gradients.
City operators who own a corridor and a clock. Fleet desks that cannot afford a 14-minute surprise. Researchers who want a graph that is not a citation. Builders who will put a camera on a pole and expect the API to speak density.
You get a live reading of which segments will fail in the next hour, which cameras already see the failure, and which detours the policy prefers. You do not get a 40-color heatmap that requires a legend.
You get the Thiruvananthapuram graph, the 5-minute cadence, the hybrid error tables, and the PPO delta against Dijkstra. The methods page is long because the work is long.
You get REST for predict, detect, influence, and route. You get GeoJSON for the map. You get the honesty that YOLO may be dark on a slim host — the rest of the loop still runs.
Historical speeds enter the LSTM. Spatial context enters the GAT. Frames enter YOLOv8. Predicted speeds become edge weights. PPO samples a path. The map draws it. A human either takes it or teaches the policy that the path was wrong. That is the product.
Predict — place speeds + PPO corridor. Map — city colors, grey direct line, shortest / fastest / balanced routes. Influence — heat map of road links near a place. Detect — YOLO counts on an upload. Network — 3D attention metaphor. Research / Technology — numbers and stack. Speeds in the interactive demo are scenario-based forecasts on real OSM geometry, labeled honestly so the product remains a lab you can trust.
The original Flask service remains the scientific runtime — preprocessing, hybrid inference, Folium maps, optional Ultralytics. This Next.js surface is the public instrument panel: the place a San Francisco design jury and a Kerala traffic engineer can both sit without either feeling talked down to.
We ship parchment and obsidian as first-class variants. The orange does not become a theme. It remains a highlighter.
Included
Interactive LSTM vs hybrid forecasts on named Thiruvananthapuram corridors, with MAE, R², and horizon in minutes.
Three.js graph of nodes and attention, plus the original OSM-derived topology of 250 / 598.
Class-wise vehicle counts as a density prior, framed like a lab still — boxes, not banners.
Which neighbors the GAT is looking at when it writes a speed.
PPO policy versus Dijkstra, reported as time saved, not distance bragged.
Predict, upload, influence, route — documented like a hardware datasheet.
Operator checklist
Name the commute that breaks — e.g. Airport → Secretariat at 09:00 in rain.
Confirm OSM roads cover that corridor. Traffic already ships a TVM drive graph.
Decide which incidents matter: rain, events, accidents, peak vs clear.
Choose whether operators want shortest, fastest, or balanced defaults.
If poles exist, Detect can turn frames into density. Privacy policy first.
Label demo vs live data in every slide. Cities forgive limits; they do not forgive silent fiction.
Common questions
No. It is a forecast-and-policy lab for corridors, fleets, and researchers — with a public instrument panel.
Yes. Prediction, map, influence, and routing work without cameras. Detection is optional density.
Thiruvananthapuram is the shipped case. The architecture is portable to other OSM drive graphs.
Product explains modules; Predict is the working studio; Developers is the contract.