Ajinkya Deshpande

Research Scientist / Engineer at Luma AI

Redwood City, CA

Resume (PDF)Email Google Scholar LinkedIn GitHub

I work on multimodal foundation models, controllable video generation, and robot learning, spanning midtraining, post-training, and inference-time guidance. Previously, I worked at Microsoft Research and Genesis Therapeutics. I earned my MS in Machine Learning from Carnegie Mellon University.

At Luma AI Jan 2026–present

Led post-training and training-free extensions for Ray3.2, enabling 10-second video generation and multi-keyframe conditioning. Developed multishot and temporally fine-grained caption conditioning for Ray.
Ran large-scale multimodal midtraining for Luma’s robotics foundation model; led sidecar ablations, integrated selected objectives, and co-designed the data and task mixture.
Led fine-tuning experiments to expand controllability beyond text prompting, owning dataset creation, training, and evaluation end to end.

Tools: Python, PyTorch, PyTorch Geometric, Hugging Face, NumPy, SciPy, scikit-learn.

Previously

Genesis Therapeutics ML Research Intern

Training-free conditioning for diffusion transformers, inference-time guidance, and sequential Monte Carlo for molecular generation.

May–Jul 2025

Carnegie Mellon · Ma Lab Research Assistant

Built MIMYR for tissue structure and gene expression, and LLOKI for spatial transcriptomics integration using GNNs and optimal transport.

Sep 2024–Dec 2025

Microsoft Research India Research Fellow

LLM agents with tool use and active RAG, RepoClassBench for code generation, and FiGURe for graph representation learning. Contributed to the second-place team in the NeurIPS 2023 LLM Efficiency Challenge.

Sep 2022–Jul 2024