Doğaç Eldenk

Eldenk studies how to make AI systems run more efficiently. His personal biography identifies work at fal on inference optimization and post-training, and earlier software engineering at Carbon Health, where he worked on developer infrastructure and scalable systems. He studied at Bilkent University and describes graduate study in computer science at Northwestern University, with research advice from Stephen Xia and Hongyang Zhang.[1][2]

An optimizer that must prove its speedups

Auto GPU Kernel gives a coding agent an optimization task and separates the search loop from measurement. The kopt tool runs the optimization agent; kbench handles correctness checks, benchmarks and experiment history. That separation addresses a recurring failure mode in automated optimization: treating noisy measurements as real improvement, or producing a faster answer that no longer meets the original specification. The repository supplies optimized kernels and a technical report.[3]

Keeping speculative decoding stable

Speculative decoding uses a smaller draft model to propose tokens that a larger model verifies. Eldenk’s site connects his work to EAGLE 3.1. The joint release explains attention drift: as draft generation continues, attention can shift toward the drafter’s own recent tokens. EAGLE 3.1 changes normalization and the hidden states passed between steps to improve stability. The release reports stronger acceptance lengths under long contexts and varying prompts, with results specific to its tested setups.[1][4]