assets/media/mirokai-demo.mp4
Drop the file in to replace this placeholder
Mirokai — Social Robot
Enchanted Tools
[TODO: caption] Mirokai social robot in interaction.
Research scientist with 10+ years bridging cognitive science and machine learning. Currently building plant genomic foundation models at Living Models. Previously multimodal AI for social robots, medical speech analysis, healthcare ML, and affective computing research.
For over a decade, I've been fascinated by one question: how do we build AI systems that truly understand humans? My journey started with a PhD in affective computing at Sorbonne University, where I developed models to recognize and synthesize social behaviors in virtual agents.
Since then, I've applied this lens to healthcare (predicting patient behavior with interpretable ML at Semeia), human-AI collaboration (modeling rapport at Inria), social robotics (multimodal perception for the Mirokai robot at Enchanted Tools), and medical speech analysis for emergency triage at e-sensia. Today I build plant genomic foundation models at Living Models.
I believe the path to beneficial AI runs through understanding the systems we model—and making AI interpretable enough that we can verify it does what we intend.
Building plant genomic foundation models across dozens of crop species. Designing long-context architectures (Mamba, SSM/transformer hybrids) at single-base resolution, and post-training beyond masked language modelling — functional supervision, regulatory atlases, and breeding-relevant readouts. Building evaluation suites meant to be meaningful to breeders rather than leaderboards, backed by an agentic research stack.
Multi-agent LLM systems for long-horizon tasks that coordinate via democratic voting and hormonal-style feedback signalling. Documented an open-loop coordination failure — signals emitted but no longer feeding back into shared state. [TODO: add repo/writeup link]
Developed multimodal speech systems for emergency triage, combining fine-tuned audio models (Whisper, Wav2Vec2, EnCodec) with raw signal processing. Led the ESYNAPSE grant — a France 2030 "Pionniers de l'IA" laureate worth up to €10M — and collaborated with clinical partners to validate models on real-world emergency medical data.
Built ML systems for the Mirokai social robot, integrating VLMs, LLMs, and speech models for natural human-robot interaction. Developed agentic pipelines with safety constraints and worked alongside a team of ML engineers.
Developed interpretable deep learning models to predict rapport in human interactions from multimodal cues. Built reusable toolkit for multimodal feature extraction and analysis.
Built predictive models on French National Health Data. Improved medication adherence prediction from 60% to 90%. Implemented SHAP/LIME for clinical interpretability. Published at NeurIPS and MICCAI.
Thesis on multimodal social signal analysis for affective virtual agents. Visiting scholar at USC ICT. Developed computational frameworks for extracting and recognizing social behaviors.
Understanding and modeling human emotions, social signals, and interpersonal dynamics through computational methods.
Making ML models transparent and explainable, from SHAP/LIME applications in healthcare to mechanistic understanding of neural networks.
Integrating vision, speech, and language understanding for robust human-AI interaction in embodied systems.
Building AI systems that interact safely and naturally with humans, with appropriate social behaviors and safety constraints.
Applying insights from human behavior modeling to build AI systems that remain aligned with human values and intentions.
Predictive models for patient behavior, treatment adherence, and care pathways with clinical interpretability.
[TODO: caption] Mirokai social robot in interaction.
[TODO: caption] 3D gesture guidance AR interface.
[TODO: caption] Ultrasound-guided prostate biopsy training simulator.
[TODO: caption] Weekly presidential addresses for stance study.