Home
FR

Loading search index…

All projects
Multimodal AIActive

Sentinelle

A multimodal voice agent that estimates cognitive load in a clinical context.

Personal project2026
  • Multimodal AI
  • RAG
  • CNN-LSTM
  • Whisper
  • TTS
  • MLOps
Abstract waveform illustration, the Sentinelle project motif

Context

Sentinelle explores an applied research question: can we estimate, continuously and without an intrusive sensor, a practitioner’s cognitive load during a clinical task?

Problem

Existing cognitive-load indicators usually need dedicated sensors that don’t fit everyday clinical use. The system needed an estimate usable with lighter multimodal signals, and a natural voice interaction that stays in the clinician’s workflow.

Architecture

  • Cognitive-load estimation: a CNN-LSTM model (MobileNetV2 + LSTM) trained on multimodal signals from several reference datasets — AVCAffe, RAVDESS, PoseTrack, BIWI.
  • RAG agent: Mistral / Ollama + LangChain + ChromaDB, to ground the copilot’s answers in reference clinical documentation.
  • Speech-to-speech voice pipeline: speech recognition (Whisper), processing, then speech synthesis (Coqui TTS), for a copilot that stays in the conversation.
  • FastAPI backend and React dashboard for monitoring, with an MLOps layer for the model lifecycle.

Stack

Python for signal processing and the CNN-LSTM models (MobileNetV2 + LSTM), Whisper for speech recognition, Coqui TTS for synthesis, Mistral/Ollama + LangChain + ChromaDB for RAG, FastAPI + React for the backend and dashboard.

Results

  • A working end-to-end speech-to-speech pipeline in a test environment.
  • A CNN-LSTM model trained and validated across four reference multimodal datasets.

Screenshots

Screenshots coming soon.