A conceptual project exploring how crowd signals and LLM reasoning can enhance public transport arrival estimation beyond conventional methods.
Most public transport tracking systems already rely on standard methods such as GPS, schedules, and historical averages.
This project does not attempt to replace those systems.
Instead, its focus is on a peculiar enhancement: leveraging crowd-derived signals and LLM-based reasoning to refine or contextualize arrival estimations.
Conventional systems answer:
“Where is the vehicle?”
This project explores:
“What does the crowd behavior suggest about the vehicle’s state?”
Examples of crowd signals:
- Sudden increase in passengers waiting
- User-reported congestion or delays
- Textual complaints, confirmations, or uncertainty
- Indirect human signals rather than sensors
The assumption is simple:
Humans collectively sense disruptions faster than systems.
Large Language Models (LLMs) can:
- Interpret noisy, unstructured crowd inputs
- Reason over uncertainty
- Produce probabilistic or qualitative updates to existing arrival estimates
This acts as a reasoning layer, not a tracking replacement.
- Assume conventional tracking already exists
- Collect simulated or hypothetical crowd signals
- Use LLM reasoning to:
- Detect anomalies
- Adjust confidence levels
- Qualitatively update ETA expectations
This repository focuses on conceptual clarity and reasoning design.
- Language: TBD
- LLMs: Conceptual / simulated
- Data: Synthetic or illustrative inputs
- Project initialized
- Core idea and differentiation documented
- Simulated crowd-input datasets
- Prompt design for uncertainty reasoning
- Lightweight prototype implementation
- Ethical considerations of crowd-sourced inference
This is an exploratory and educational project. It is not intended for production deployment.
-
A mock crowd-signal file (
sample_crowd_signals.json) represents simplified human-derived observations such as complaints, confirmations, and crowd size. -
The Python module reads this input and applies a rule-based reasoning layer that mimics how an LLM might qualitatively assess uncertainty.
-
The output is not a prediction, but a confidence-oriented assessment (on-time / delayed / uncertain).
This keeps the project transparent, explainable, and extensible.
Most transit systems rely on sensors and schedules.
This project explores how human crowd signals, interpreted through reasoning systems, could act as a complementary layer—especially when traditional signals are weak or delayed.
Crowd-derived signals are inherently noisy and incomplete. A single complaint or confirmation does not represent ground truth.
For this reason, the system does not attempt to output precise predictions or exact arrival times.
Instead, it produces confidence-oriented assessments, such as:
- High confidence: likely on time
- Medium confidence: uncertain state
- Low confidence: possible delay
This mirrors how a reasoning system (or LLM) would operate: by synthesizing imperfect signals and expressing degrees of belief, not absolute certainty.
In real-world transit systems:
- Human signals arrive earlier than sensor updates
- Data may be contradictory
- Overconfidence can be misleading -Qualitative confidence bands are safer, more interpretable, and easier to integrate as a supplementary layer to existing tracking systems.
This project prioritizes:
- Explainability over prediction accuracy
- Confidence adjustment over hard outputs
- Human signals as complementary inputs
The intent is to explore reasoning patterns, not to replace existing production systems.