QUT

Mobility Intelligence

Checking model…
Live mobility mapModel predictions and grounded recommendations
D DestinationE ChargersP Places
Configure the browser key and run a prediction.
TRAINING PIPELINE

Training data

What the mobility model learns from, how a trip becomes model input, and which metadata is stored in the deployed checkpoint.

Checkpoint-backed
Loading dataset metadata…

Dataset views used in the notebook

Raw recordsOriginal Porto taxi rowsTrip identifiers, call metadata, timestamps, missing-data flag, and encoded POLYLINE coordinates.
Cleaned trajectory eventsPrediction checkpointsValidated trips with observed prefixes, movement summaries, time context, and destination labels.
Model-ready tensorsInputs used for trainingScaled context vectors, 32×4 trajectory sequences, and destination-cluster targets, displayed as latitude and longitude.

Dataset partitions

Counts appear when they were exported with the checkpoint.

Model-ready input

Context features

Trajectory preparation

1Raw POLYLINEGPS coordinates
→
2Observed prefixNo future points
→
3Resampling32 trajectory states
→
4Relative motion4 sequence features
→
5ScalingTraining scalers
DEPLOYED CHECKPOINT

Model structure

The selected architecture, tensor flow, layers, and stored validation evidence.

Loading…
Loading model metadata…

Selected model flow

Layer structure

Validation metrics

Architectures evaluated

MLP18 context featuresDense baseline for engineered trip context.
LSTM32 × 4 sequence + contextModels ordered movement before classification.
GNN32-node trajectory graph + contextModels neighboring trajectory-state relationships.
GROUNDED MOBILITY INTELLIGENCE

How the LLM and agents work

The LLM explains verified outputs; it does not predict the taxi destination or invent live place evidence.

Evidence grounded
SYSTEM ARCHITECTURE

Prediction, tools, LLM, and interface

Each layer has a separate responsibility. Verified outputs move forward through the pipeline and progressively update the live planner.

ML predictionLive toolsLLM reasoningUser interface
Input layer
01
User journeyCall type · timestamp · observed GPS prefix
02
Mobility requestQuestion · categories · shared search radius
Feature preparation→
Prediction layer
ML
Selected trained modelMLP loaded from the supplied checkpoint
Destination point + confidenceThree most likely destination points (D1–D3) with confidence
Destination D→
Mobility-service layer
G
Google PlacesUseful places and EV charging stations around D
R
Google RoutesDriving distance and duration from D
Verified evidence→
Agent layer
A
Workflow orchestratorRuns prediction, searches, routing, ranking, and live events
LLM
Grounded LLMExplains and organizes only the supplied evidence
Streaming result→
Experience layer
MAP
Interactive mapRoute · D · alternatives · E1–E8 · P1–P5
UI
Live mobility planActivities · probabilities · ranked evidence · final answer
Grounding boundaryThe trained model predicts the destination. Google services supply current place and route evidence. The LLM may summarize and recommend from those inputs, but it cannot replace or modify them.

What the model does

  • Transforms the observed trip prefix into context and sequence tensors.
  • Scores candidate destination locations across Porto.
  • Ranks the three most likely next-destination points (D1, D2, D3) with confidence.

What the agent does

  • Searches only around D.
  • Combines destination confidence, rating, and route evidence.
  • Explains a primary place, charging option, alternatives, and uncertainty.

What external services do

  • Google Places returns venues and EV charging stations.
  • Google Routes provides driving distance and duration.
  • Google Maps renders the trajectory and result markers.

Grounding safeguards

  • The prompt forbids invented places, ratings, availability, and distances.
  • Missing evidence is displayed as unavailable.
  • OpenAI and server-side Google keys never enter browser JavaScript.
INTERACTIVE LEARNING

Neural network simulation

Explore how inputs, weights, activations, loss, and parameter updates work during training.

INTERACTIVE LEARNING

Transformer simulation

Follow tokenization, embeddings, query-key-value attention, multi-head attention, and output generation.

INTERACTIVE LEARNING

Agent loop simulation

Observe how an agent receives a goal, plans, acts through tools, evaluates evidence, and revises its next step.