CV-MAPS: An Interactive Multi-Agent System for Personality-Aware Multimodal Self-Presentation Coaching

This repository contains the demo system described in the AAMAS 2026 paper "CV-MAPS: An Interactive Multi-Agent System for Personality-Aware Multimodal Self-Presentation Coaching".

Authors: Elena Ryumina, Alexandr Axyonov, Dmitry Ryumin, Alexey Karpov.

Abstract

This demo introduces CV-MAPS, an interactive multi-agent system for personality-aware multimodal self-presentation coaching. The system guides users through recording short multimodal CVs, which are analyzed by a neural network to estimate emotions and personality traits from face, body, voice, scene, and transcribed speech. CV-MAPS leverages off-the-shelf Large Language Models (LLMs) to generate personalized, actionable feedback on speech content, emotional regulation, nonverbal, and para-linguistic presentation. Users can iteratively and interactively refine their multimodal self-presentations, while the system tracks progress and suggests alternative professions when alignment with the target profile remains low. By closing the loop between multimodal expression, AI analysis, and coaching, CV-MAPS enables data-driven profession exploration through authentic self-presentation.

Demo video: https://youtu.be/Z41Hpe_OROQ

System Overview

CV-MAPS implements a closed-loop process with five coordinated agents:

• Profession Guidance Agent: manages UI and session flow.
• Recommendation Generation Agent: drafts interview questions for the target profession.
• Multimodal Personality and Emotion Estimator: analyzes face, body, scene, audio, and text.
• Feedback and Coaching Agent: interprets analysis and provides actionable advice.
• Progress Tracking Agent: compares attempts, visualizes dynamics, and suggests alternative roles.

How It Works

1. Select language and target profession.
2. Generate interview questions.
3. Record or upload a video response.
4. Run multimodal analysis (emotion + Big Five traits).
5. Receive personalized feedback and suggested improvements.
6. Record additional attempts and track progress toward the target profile.

Project Structure (high level)

• app/ UI and orchestration logic
• core/ models, inference, attribution, metrics
• data/ profession trait profiles
• assets/ custom CSS
• run.py entrypoint

Configuration

Runtime and UI settings live in app/config.py (e.g., thresholds, limits, language defaults, and feature flags). Adjusting this file lets you tune the demo without changing the core logic.

Setup

1. Create and activate a virtual environment.
2. Install dependencies:

   pip install -r requirements.txt

3. Download the core/modalities folder from Google Drive and place it into core/ with replacement: https://drive.google.com/drive/folders/1ccUYUFJAbqGlt4atFK4h43MT1lkGx2JY?usp=sharing
4. Ensure ffmpeg is available in your PATH (used for video conversion).

Notes:

• GPU is recommended. The demo auto-selects CUDA if available.
• The first LLM run will download the Qwen model into the local cache.

Run

python run.py

Open http://0.0.0.0:7860 in a browser.

Acknowledgments

We thank the authors of the study below for publishing the profession-trait profiles used in this demo:

@article{kern2019social, title={Social media-predicted personality traits and values can help match people to their ideal jobs}, author={Kern, Margaret L and McCarthy, Paul X and Chakrabarty, Deepanjan and Rizoiu, Marian-Andrei}, journal={National Academy of Sciences}, volume={116}, number={52}, pages={26459--26464}, year={2019}, }

Data source: https://github.com/behavioral-ds/VocationMap/blob/master/data/profession-profiles.csv