Gregory Axler

Experience

• Symbolica AI | Founding Research Engineer | Apr 2023 – Present
  • Developed efficient deep generative models for reinforcement learning-based fine-tuning of LLMs, including methods for synthetic data generation.
  • Full lifecycle ownership including data pipelines, distributed training, and model architecture.
  • Contributed to team recruitment and engineering best practices.
• AI Teach U | Lead Research Engineer | Nov 2018 – Dec 2023
  • Designed and built deep learning models for computer game animation, including video-to-3D facial animation system and audio synthesis using GANs, optical flow, pose estimation, and flow-based models.
  • Developed infrastructure for automated data ingestion and model retraining.
  • Managed full ML pipeline from data processing to deployment.
• ClearML (formerly Allegro.AI) | Research Scientist | Nov 2016 – Dec 2019
  • Developed and implemented novel solutions for diverse computer vision research projects, including CNN architecture optimization, self-improving object detection, semantic segmentation, and video search engine.
  • This included building end-to-end training and evaluation pipelines.

---

Projects

• Efficient Theorem Proving via Structured Generative Models (Symbolica)

  • Developed a structured generative model, theorem identification algorithm, and synthetic data generation methods to enhance RL-based theorem proof search in a distributed environment.
  • Achieved a 2.5x reduction in generation steps and a 50% faster search time.
  • Addressed GPU underutilization and combinatorial complexity of theorem identification.

• Llama Inference Engine in Pure Rust (Personal Project)

  • Built a high-performance Llama inference engine from scratch in pure Rust (github repo).
  • Optimized matrix-vector multiplication, parallelized computation, and implemented quantization.
  • Demonstrates expertise in transformer models, Rust programming, and performance optimization.

• OpenAI Triton Compiler Contributions and Rust-CUDA Transpiler Development (Symbolica/Personal)

  • Developed high-performance GPU algorithms using CUDA and Rust, including a CUDA-based compression algorithm that reduced memory needs by 98%.
  • Contributed to OpenAI's Triton compiler and developed a Rust-CUDA/Triton transpiler at Symbolica.
  • Enabled simple integration of GPU kernels inside Rust code at Symbolica.

• Optimized CNN Architectures for Object Detection on Edge Devices (ClearML)

  • Designed and implemented optimized CNN architectures for embedded systems, achieving a 150% inference speed-up through dynamic network sizing.
  • Developed a technique that adjusts the network size based on the predicted complexity of a frame (resulting in a US patent).

---

Education

• MSc Operations Research (Magna Cum Laude) | Tel-Aviv University | 2018
  • Publication: "Toward a Dataset-Agnostic Word Segmentation Method" | IEEE ICIP
  (Project Page)

---

Skills & Tools

• ML/DL & Analysis Tools: Torch, JAX, Triton, Pallas, HuggingFace, Ray, Numba, OpenCV, Pandas/Polars, Seaborn, SQL
• Programming Languages: Python, Rust, CUDA, C, C++ (Some: Haskell, Javascript, Go)
• Infra & Serving: FastAPI, Docker, Kubernetes, AWS, GCP, W&B