Daniel Fleischer
Principal Software Engineer, AI Frameworks at Microsoft
I am a Principal Software Engineer in Microsoft’s AI Frameworks group, which works on running AI models efficiently in Microsoft’s data centers. Before that, I was at Intel, working across the LLM stack: agentic systems, reinforcement learning for reasoning, retrieval-augmented generation, fine-tuning and inference optimization. My focus there was the transition from research to production — systems that are not only demonstrated, but deployed and measured.
Much of that work is open source. Xe-Forge is an agentic system that autonomously optimizes GPU compute kernels, combining a planner and tool-runner architecture with a curated knowledge base of Intel GPU expertise that is largely absent from model training data. DeepMath is a reinforcement-learning framework for training tool-using math agents with GRPO, where the model writes and executes Python within its own reasoning. fastRAG and RAG-FiT address efficient retrieval-augmented generation and the fine-tuning of language models for RAG tasks, and Dicta-LM 2.0 is an open Hebrew language model pretrained on Intel Gaudi accelerators.
A consistent theme across these projects is measurement. Agentic systems are straightforward to demonstrate and difficult to trust, and my work therefore tends to include the evaluation harness and benchmarking infrastructure alongside the model itself.
My background is in physics: an M.Sc. on dualities in supersymmetric field theories and doctoral research in astrophysics, both at the Weizmann Institute of Science. I am also a long-time Emacs user and a contributor to the GNU Emacs project and ecosystem.