Compiling the hour
Technology
We did not invent a new kind of car. We invented a new kind of honesty about the next hour of road. The stack is classical on purpose: memory, attention, vision, policy — each named, each measured, none hiding behind a slogan.
Temporal memory
A road is a time series before it is a place. The LSTM is trained on 5-minute samples across 60 days. It learns the grammar of peak: the Monday that is not a Friday, the rain that is not a holiday. Standalone, it posts MAE 0.5354, RMSE 0.7880, R² 0.9766 — a strong univariable singer.
Spatial attention
Graph Attention refuses the fiction that an edge is alone. Each node queries its neighbors and writes an attention coefficient. Congestion becomes a rumor with a source. The hybrid LSTM–GAT posts MAE 1.3447 and R² 0.9309 — numerically louder error, spatially better truth. Routing needs the second number.
Density from pixels
Loop detectors lie by omission. A camera does not. YOLOv8n detects bicycle, car, motorcycle, bus, truck. Counts become a feature, not a screenshot. On slim hosts the detector can sleep; the rest of the stack does not wait for it.
Adaptive policy
Predicted speeds become living weights. Proximal Policy Optimization learns a routing policy that is allowed to look foolish on a paper map if it is wise on a stopwatch. Versus Dijkstra: 10–25% less travel time, smoother load, fewer heroic U-turns.
Architecture
Data arrives as OSM topology via OSMnx and NetworkX, and as a 60-day multivariate stream: speed, flow, congestion, weather, density. Preprocessing is unromantic and non-negotiable — missing values, outlier interpolation, time sync, min-max, graph index mapping.
The LSTM consumes the temporal window. Its hidden state is the city’s short-term memory. The GAT consumes the same moment as a graph: each road asking its neighbors how much of their present should become its future. Fusion is the hybrid forecast. That tensor is not a chart. It is an edge-weight field.
Dynamic weights enter the PPO environment. The agent proposes a path. Reward is the negative of experienced time, with penalties for re-entering a predicted jam. The clip in PPO is the adult in the room: updates stay proximal so a noisy afternoon cannot unteach a month.
A univariate LSTM can hug a single series until the residual is tiny and the route is still stupid. GAT spends capacity on neighbors. That shows up as MAE you can quote against us, and as detours you can actually drive. We print both numbers. The Ember CTA is the routing one.
PyTorch for the nets. Ultralytics for detection. Stable-Baselines3 for PPO. Folium for the scientific map. Three.js on this site for the public graph — a cousin of the adjacency matrix, not a replacement. Flask remains the lab server. Next.js 16 is the instrument face.
Nothing in this stack is a secret sauce. The sauce is the closed loop and the refusal to ship a predictor that cannot move a vehicle.
Five-minute bins, because finer is vanity without sensors. A real capital graph beats a synthetic grid. Optional YOLO, because not every host should download weights to say hello. Ember only on actions, because a technology page that shouts has already lost the argument.
Predict and Map consume demo edge speeds + OSM geometry + place search (Nominatim / Overpass catalog). Influence builds a local attention-style matrix from those speeds near a place. Detect calls YOLOv8 when weights are present. Network is a Three.js teaching graph. Keeping the stack named this way stops “AI traffic” from becoming a single vague button.
Data path
01
OSMnx extracts Thiruvananthapuram. Drive graph, weather, density, 5-minute stride.
02
Interpolate, sync, normalize, map nodes to graph indices. No silent NaNs into the LSTM.
03
LSTM memory + GAT attention (+ YOLO density) write future speeds onto edges.
04
PPO reads weights, emits a route, Folium and Three.js both tell the truth.
Glossary
A directed piece of road between intersections in the OSM graph.
How much a road listens to a neighbor when writing a forecast.
How far ahead you ask the model to speak, usually in 5-minute steps.
A learned chooser of next roads that optimizes time under predicted speeds.
How many vehicles a camera (or estimate) sees — a feature, not a fine.
Lightweight speed field used in this deployment so demos run without a GPU.
Common questions
LSTM is memory of time. GAT is memory of neighbors. Naming both stops “AI” from becoming one blurry box.
On time, research showed 10–25% gains in the evaluated setting. Distance-only Dijkstra can still win if you only care about kilometres.
Flask scientific runtime and Next.js UI both live under traffic-app. Developers page lists the HTTP surface.