Teaching artifical intelligent models to be intelligent :)

I am passionate about Artificial Intelligence and Machine Learning research. My research interests span agentic ai, ai for science, computer vision, representation learning, generative models, multimodal learning, and responsible AI. I am driven by the potential of artificial intelligence (AI) to address complex challenges and create impactful solutions. :)

Outside of work, I love traveling to explore new countries and cultures. I'm also deeply committed to education, and I have volunteered in various countries to support those without access to learning opportunities.

identity:
  origin: China 
  nationality: France
  tagline: Engineering the intelligence in AI & Automating with it
  travel_into: [Singapore, Thailand, Malaysia, Japan, China, Taiwan, 
               France, Spain, Italy, Germany, Sweden]

passions:
  culture: 🌍 Traveling & Exploring Cultures
  activity: 🥟 Gyoza Making & Eating
  sport: 💃 K-Pop Dancing
  love: ❤️ Girlfriend Caring
  gaming: 💥 Counter Strike

philosophy:
  ego_check: "Low ego, high impact."
  success: "Follow what you like, success will follow."
  open_source: "Open source is magical."
  resilience: "Failure is a form of success." 

🌱 ourtub: seeing knowledge, not just searching for it

🌱 jaxlaxy: your compass for the jax multiverse

🌱 sweepkit: rust-powered blazing-fast cli tool to scan & clean unused dev dependencies

🌱 torchOCR: pytorch library for end-to-end optical character recognition

🌱 courant: cozy reminders for developers to drink water, rest eyes, and stretch

🌱 rubiksera: learn to solve your rubik's cube

🌱 mantispdf: free pdf tools in your browser

🔖 huggingscience: partial differential equations (pde) importance

🔖 responsible ai: impact of knowledge distillation on model interpretability

🔖 ml4sci: physics-guided machine learning on gravitational lensing

🔖 ott-jax: implementation of annealed Sinkhorn for optimal transport (convergence, regularization path and debiasing)

🔖 graph generation: neural graph generation with text conditioning

🔖 listen to the wild: predicting naturalness with acoustic indices

🔖 imagenet-sketch classification: classify b&w google image sketch of standard imagenet class

🔖 vlm few-shot performance improvement: enhance sparse attention selection mechanism

🔖 composed video retrieval: improved alignment between visual and textual embeddings for composed video retrieval

🔖 hygene: diffusion-based hypergraph generation method

🔖 cfdg: review of classifier-free diffusion guidance

🔖 unifying gan & diffusion: score gan and discriminator flow, a unified framework

🔖 tiger: generative retrieval of item IDs for recommender systems

more?

💼 Experiences

Role	Description
Lead AI Engineer @ (December 03, 2025)	Managed a 3-person team to develop an aviation maintenance document processing platform using agentic AI workflows (from 0/ideas to $1M).
Research Data Scientist @ (February 24, 2025)	Generated coverage maps using an image segmentation model in the GeoAI domain, automated anomaly detection (Prophet & SARIMAX), built ETL data pipelines, and created insightful dashboards for directors.
Research Engineer @ (December 12, 2024)	Enhanced sparse attention selection mechanisms to boost few-shot classification performance on Vision Language Models. (See Report)
Research Engineer @ (November 5, 2024)	Improved alignment between visual and textual embeddings for composed video retrieval by designing a novel loss function and an MLP architecture on the CoVR research paper. (See Paper)
Research Engineer @ (March 2, 2024)	Created an efficient checkpointing fine-tuning scheme for Deep Neural Networks (DNNs) using Delta-LoRA + LC-checkpoint, achieving compression ratios up to 25x on models like ViTs, ResNets, VGGs, AlexNet, and LeNet. (See Code)
Research Engineer @ (June 8, 2023)	Developed an interactive optimization algorithm for a Constraint Satisfaction Problem (CSP), applying Neural Networks and Decision Trees and engineering techniques to improve IBM's CPLEX solution generation. (See Code)

---

🗂️ Projects

Youtube Comments Sentiment Analysis	Retrieval-Augmented Generation (RAG) System	HYGENE: Diffusion-based Hypergraph Generation
A full MLOps pipeline using Docker, GitHub Actions, DVC, and MLflow.	A RAG system leveraging LangChain, FAISS, and Ollama for inference.	Implemented hypergraph generation with diffusion models (AAAI), a project at Télécom Paris.
Groundwater Level Prediction	Neural Graph Generation from Text	Classifier-Free Diffusion Guidance (NeurIPS)
A project with Hi!Paris & Inria to predict groundwater levels, which received an Honorable Mention.	Developed a Neural Graph Generator to create complex graph structures from text at École Polytechnique.	A paper implementation focused on jointly training conditional/unconditional diffusion models.
ImageNet-sketch Classification	EuroSAT Image Classification	Predicting Naturalness with Acoustic Indices
Classifying sketch images with EVA-CLIP finetuned with an extra-MLP layer.	Classify geospatial images with ResNet from the EuroSAT dataset.	Used scikit-maad and VGGish to extract acoustic indices and apply ML models to find patterns and predict naturalness.
Presentation: Generative Recommender Systems	Presentation: Unifying GANs & Diffusion
Presented TIGER, a method for generative retrieval of item IDs for recommender systems.	Presented Score GAN & Discriminator Flow, a unified framework for GANs and Diffusion models.

---

🎖️ Honors

Mean Arterial Pressure Prediction	Full-stack Agent-Agnostic Solution	Skill-Gap Analysis Automation
Rank: 1st / 200+	Rank: 2nd / 400+	Rank: 2nd / 650+
Achieved 1st place at the Inria challenge with a domain adaptation-aware model for Mean Arterial Pressure (MAP) prediction.	Placed 2nd at the MIT Hackathon with a full-stack solution (LangGraph, OpenAI, FastAPI, Next.js) to trace and visualize AI agent interactions.	Placed 2nd at the MIT Hackathon with a full-stack app (React, FastAPI, Gemini, BeautifulSoup) to automate skill-gap analysis by aggregating multi-source data.

---

🌎 Open Source

PyTorch / torchtitan	Hugging Face Transformers	OTT-JAX	Responsible AI Blog (IP Paris)
Added missing --model.flavor=debugmodel_flex_attn flags for Llama 3 FlexAttention in the compiler toolkit documentation.	Updated model cards for the Swin Transformer, Swin V2, Pixtral & ShieldGemma 2 models to improve documentation.	Authored a tutorial on Annealed Sinkhorn for the Optimal Transport Tools (OTT-JAX) library.	Wrote a blog post analyzing the impact of knowledge distillation on model interpretability.

---

🎓 Visiting​

Institution	Focus
Harvard University	Completed renowned courses providing a robust foundation in Computer Science and AI (CS50x, CS50P & CS50AI).
Stanford University	Followed world-class courses in Computer Science and Artificial Intelligence.

---

I'm always excited to take on new challenges in AI research and application. If you have an interesting project, a research idea, or just want to discuss the latest in tech, let’s connect! I'm open to collaborations and geeking out about all things AI :)