prof_pic.jpg

Doğaç Eldenk

Evanston, IL
[email protected]

I am a Machine Learning Engineer / Researcher at fal.ai, working on inference optimization and post-training methods on generative models.

I am also a Master’s student at Northwestern University, studying Computer Science, graduating in August 2026. I’m advised by Prof. Stephen Xia (Northwestern University) and Prof. Hongyang Zhang (University of Waterloo). My research focus is mainly AI and ML systems efficiency. I’ve mainly worked on speeding up LLM inference using speculative decoding, training new models and improving architecture, ultimately releasing EAGLE 3.1 with vLLM and LightSeek Foundation.

Previously I’ve worked at Carbon Health as a Software Engineer for over 4 years, mostly as a platform engineer to improve developer efficiency and build scalable systems. I am a big fan of open-source and I occasionally write blog posts about topics that I find interesting, mostly related to software.

news

Jan 23, 2026 Blog moved to the personal website, from Ghost to Jekyll

latest posts

Apr 03, 2026 Setting up Ephemeral GPU workspaces in Modal for SGLang Development
Feb 03, 2026 Training language models on TPUs shouldn't be scary
Jun 11, 2025 The Case Against Dependency Injection

awards

  1. 2x🥇 Winner – MLSys 2026, NVIDIA FlashInfer GPU Kernel Writing Competition
    Won 1st place twice in DeepSeek Sparse Attention for AI assisted and AI-only tracks.

publications

  1. attentiondrift.png
    Attention Drift: What Autoregressive Speculative Decoding Models Learn
    Doğaç Eldenk, Payal Mohapatra, Yigitcan Comlek, and 3 more authors
    2026
    PDF GitHub stars
  2. modalityaware.png
    Modality-Aware Zero-Shot Pruning and Sparse Attention for Efficient Multimodal Edge Inference
    Yueyuan Sui, Payal Mohapatra, Doğaç Eldenk, and 5 more authors
    2026
    PDF
  3. uniference.png
    UNIFERENCE: A Discrete Event Simulation Framework for Developing Distributed AI Models
    Doğaç Eldenk and Stephen Xia
    In International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things (FMSys), 2026
    PDF GitHub stars
  4. earsleeve.png
    EarSleeve: Transforming Everyday Earphones into a 12-Lead ECG Sensing Platform
    Junxi Xia, Doğaç Eldenk, Hongjun Xu, and 2 more authors
    In International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys), 2026
    GitHub stars
  5. microservice.png
    Incidents During Microservice Decomposition: A Case Study
    Doğaç Eldenk and H. Alperen Çetin
    In International Conference on Evaluation and Assessment in Software Engineering (EASE), 2025
    Best Paper Award PDF
  6. unsup.png
    Learning portrait drawing with unsupervised parts
    Burak Tasdemir, Mustafa Goktan Gudukbay, Doğaç Eldenk, and 2 more authors
    International Journal of Computer Vision, 2024
    PDF

selected projects

  1. EAGLE 3.1
    EAGLE 3.1
    Faster & resillient speculative decoding for LLM inference acceleration. Released fastest speculators for Kimi K2.6 and GPT-oss 20B/120B with _TokenSpeed_ team.
    Hugging Face Downloads
  2. Auto GPU Kernel
    Auto GPU Kernel
    Autonomous GPU-kernel discovery & optimizer, Winner 🏆 (Agent-only) MLSys 2026 - NVIDIA / FlashInfer AI Kernel Generation Contest
    GitHub stars
  3. Universal DeepSeek OCR 2
    Universal DeepSeek OCR 2
    Updated version of DeepSeek-OCR and DeepSeek-OCR 2 models to support non-CUDA backends, MPS and CPU. Downloaded over 20K+ times.
    Hugging Face Downloads
  4. Defenchess – AI Chess Engine
    Defenchess – AI Chess Engine
    A high-performance C++ engine for advanced chess analysis and play. Previously ranked in the top-10 worldwide with a 3400+ ELO performance.
    GitHub stars
  5. kotlinx-protobuf-gen
    kotlinx-protobuf-gen
    Generate kotlin data classes from protobuf files that supports Kotlin Native that can be serialized and deserialized to protobuf using `kotlinx.serialization`.
    GitHub stars