Experience

🐡 Sakana AI

Research Scientist - Open-endedness

Tokyo, Japan
10/2024-Present

• Open-Ended Model and Synthetic Task Coevolution: Implement system for endlessly coevolving an LLM-based multi-agent system alongside a continuously evolving synthetic task pool.

  • Model Archive Evolution: Leverage evolutionary model merging as genetic operators and quality-diversity algorithms for selection. Our population achieves better coverage than baselines on downstream benchmarks.

  • Task Archive Evolution: Design LLM-based agent to generate novel synthetic tasks implemented in code.

  • Infrastructure: Implement system using Celery orchestration across GPU workers, handling model merging, real-time vLLM inference, and sandboxed Docker evaluation with seamless scalability.

• Diffusion for LM reasoning: Distill the reasoning process of an LLM into a Diffusion Model.

• VLM Webagents: Develop Process Reward Model (PRM) for training VLM-based Webagents with RL.

📑 TU Darmstadt

Student Researcher - Reliable Multimodal AI Lab (Prof. Marcus Rohrbach)

Darmstadt, Germany
10/2023-09/2024

• Modelling: Develop multimodal LLM (MLLM) for video understanding using PyTorch. Finetune image-text pretrained MLLM on video moment localization data, leveraging distributed training across 8 A100 GPUs.

• Results: Beat state-of-the-art models by up to 11% on 3 benchmarks. First-author paper Chrono.

🔬 Fraunhofer Heinrich Hertz Institute

Student Researcher - Applied Machine Learning

Berlin, Germany
04/2022-09/2023

• Data Engineering: Implement scraping and preprocessing units, retrieving data from varying sources.

• Neural Network Modelling: Develop, implement, and tune graph neural network architecture for time series forecasting of dynamic

  traffic flow using PyTorch.

🔬 Fraunhofer Heinrich Hertz Institute

Student Researcher - Photonic Components

Berlin, Germany
02/2019-03/2022