MAPA: Accessibility Profiling and Personalized Journey Planning

MAPA (Multimodal multi-Agent Profiling for Accessibility) is the prototype developed for the thesis Accessibility Profiling and Personalized Journey Planning: A Multimodal Multi-Agent LLM-Based Framework.

The system asks the user for permission, collects practical needs for a route, such as step-free access, spoken or written guidance, simple wording, and landmark support, and stores them as a checked JSON profile. The confirmed profile is then used to adapt fixed route examples and add selected City of Zurich open data. This repository contains the code and evaluation files used for thesis review.

Live Demo

The frontend is deployed at https://mapaiprofileragent.vercel.app.

The public demo runs in Mock mode by default, so it can be tried without a local model server or backend. Backend mode is intended for local testing with the Python API.

System Overview

Thesis Review Version

The reviewed thesis version should be cited from a fixed Git tag or GitHub release. After the submission deadline, keep that tagged version unchanged. If development continues after submission, use a new branch, a new release tag, or a separate repository.

Suggested release tag for the submitted version:

git tag -a thesis-submission-2026-07-01 -m "Thesis submission review version"
git push origin thesis-submission-2026-07-01

Repository Contents

Path	Purpose
backend/	API, profile and planning pipelines, providers, schemas, and evaluation code
frontend/	Browser interface for consent, profiling, route planning, speech controls, and map display
results/	Generated evaluation reports, case-level CSV files, and LaTeX table fragments used in the thesis
docs/	API contract, implementation notes, Zurich open-data integration report, and README assets
skills/	Agent responsibility notes used during staged implementation
pipeline_workflow.html	Interactive overview of the profile and planning pipelines

Main Features

• short profiling dialogue with permission, skip, and confirmation steps
• validated accessibility_profile JSON representation for vision, hearing, mobility, and cognition needs
• deterministic mock mode for offline evaluation
• optional local text and vision model providers
• route planning over fixed route fixtures: route_with_stairs, step_free_route, and long_walk_route
• City of Zurich open-data enrichment for accessibility barriers, toilets, and parking where coordinates are available
• REST and Server-Sent Events endpoints for profile and planning calls
• browser frontend with Mock, local model, and Backend modes

Quick Start

Create the Python environment and install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run the backend:

uvicorn backend.app.api:app --reload --port 8000

Run the frontend:

cd frontend
npm install
npm run dev

The default mode is deterministic and offline. Local model calls are optional.

API Endpoints

Method	Path	Description
GET	/api/health	Health check
GET	/api/routes	List route fixtures
GET	/api/zurich/data	Fetch Zurich barrier, toilet, and parking data near a point
POST	/api/audio/transcribe	Upload microphone audio for transcription
POST	/api/profile/turn	Profile dialogue turn
POST	/api/profile/stream	Streaming profile dialogue turn
POST	/api/plan	Create a personalized route plan
POST	/api/plan/stream	Streaming plan generation with per-node progress

See docs/API_CONTRACT.md for request and response shapes.

Evaluation Artifacts

The thesis evaluation artifacts are stored in results/.

File	Content
experiment_report.json	Full generated report for profiling, planning, boundary, and indoor-style adaptation experiments
profiling_cases.csv	Case-level profiling outputs and label scores
planning_cases.csv	Case-level planning outputs across baseline and profile-conditioned settings
boundary_planning_cases.csv	Boundary cases for fallback warnings and unsupported accessibility claims
indoor_adaptation_cases.csv	Indoor-style persona adaptation cases
table_*.tex	LaTeX table fragments included in the thesis

Run the evaluation harness:

python -m backend.app.evaluation.run_eval

Run the test suite:

pytest -q

Local Models

The mock providers are sufficient for deterministic review. To try the local model mode, start a local model server and use the model names configured in the backend provider settings.

ollama serve
ollama pull qwen3.5:4b
ollama pull llava:7b

Citation

Use the repository citation metadata in CITATION.cff, or cite the review release once the thesis submission tag has been pushed.

License

This repository is released for thesis review and academic non-commercial reuse with attribution. See LICENSE for the full terms.