Built for: National Road Safety Hackathon 2025, IIT Madras
Team: [MechaSys]
Routesit is a locally operable AI tool designed to assist engineers and road authorities in planning, prioritizing, and simulating road safety interventions. Unlike a simple GPT bot, Routesit:
- Integrates text input, optional images, and road metadata
- Retrieves and reasons over an expanded, evidence-backed intervention dataset
- Optimizes interventions based on cost, impact, dependencies, and conflicts
- Produces actionable, field-ready recommendations with citations
- Visualizes scenario comparisons for informed decision-making
Routesit is designed to be offline, lightweight, and demonstration-ready, highlighting technical depth and practical feasibility for real-world road safety planning.
Input Types:
- Free form text describing road safety issues
- Optional images for visual analysis (e.g., faded signs or markings)
- Metadata: road type, speed limit, traffic volume, accident history
Processing Pipeline:
- Data Interpretation & Normalization — cleans input, parses metadata, optional image analysis (YOLOv8)
- Intelligent Retrieval — searches vectorized intervention database (FAISS/Chroma embeddings)
- Local Reasoning — LLM ranks interventions, predicts impact, flags dependencies/conflicts
- Scenario Optimization — compares multiple intervention combinations (cost/impact/urgency)
- Output Generation — produces interactive reports, charts, step-by-step instructions, and citations
- Language: Python
- Embeddings & Retrieval: sentence-transformers, FAISS/Chroma
- LLM Reasoning: Mistral 7B GGUFA 4bit - 10 bit (local, quantized)
- Computer Vision (optionally under development): YOLOv8 for sign and road feature detection
- Dependency/Conflict Graph: NetworkX
- Visualization: matplotlib / plotly
- Frontend Demo: Streamlit / Flask
- Not Fully Classified (Under Development)
- Base: Hackathon-provided 50 interventions
- Expanded: Additional interventions curated from IRC, MoRTH, WHO, and global best practices etc.
- Annotated with: dependencies, cost brackets, predicted impact, implementation complexity, and references
Quality > quantity, each intervention entry is actionable, referenced, and ready for scenario simulation.
- Multi-modal Input Fusion: Text + photo + metadata interpretation
- Dependency & Conflict Reasoning: Ensures interventions are compatible and feasible
- Scenario Simulation: Compare “A+B” vs “C+D” to predict crash reduction and cost-effectiveness
- Actionable Output: Generates field-ready recommendations with justification and citations
- Offline Operation: Runs locally on modest hardware for hackathon demo
- IRC:67-2022, Clause 14.4
- MoRTH Guidelines
- WHO Road Safety Reports
- Public domain international best practices
Members: Divine R | Anand S
Email: [mechainthemail@gmail.com]
Hackathon Submission: National Road Safety Hackathon 2025, IIT Madras