Artificial Intelligence Researcher

Artificial Intelligence Researcher — Helix Cognitive Systems

San Francisco, CA | July 2023 – Present

• Co-developed an 8B parameter Large Language Model, improving downstream benchmark accuracy by 14% across 8 reasoning evaluations.
• Optimized PyTorch distributed training pipelines using Megatron-LM and Deepspeed, reducing training time for foundation models by 18%.
• Designed custom CUDA kernels in Triton for sparse attention mechanisms, decreasing GPU memory utilization by 22% during training.
• Authored 2 papers accepted at NeurIPS and ICML on efficient fine-tuning techniques, representing the company at international panel presentations.

AI Research Engineer — Aetheria Labs

Seattle, WA | September 2021 – June 2023

• Engineered a retrieval-augmented generation (RAG) framework that reduced model response hallucination rates from 18% to 4%.
• Trained and deployed a multimodal computer vision model for automated visual inspection, achieving a 99.4% F1-score in production.
• Refactored data preprocessing pipelines with Ray, scaling dataset loading speed by 35% for images and audio files.
• Collaborated with product teams to transition research prototypes into API endpoints, supporting 1.2M daily active user requests.

Deep Learning, Transformers, Reinforcement Learning (RLHF), LLM Pre-training, Parameter-Efficient Fine-Tuning (PEFT), PyTorch, JAX, Triton, DeepSpeed, Ray, Hugging Face, CUDA, Distributed Training, Python, C++, MLOps, Git, Linux

Open-Source Sparse Attention Kernel

Role: Creator & Core Maintainer

Tools: Triton, Python, CUDA, PyTorch

Developed a custom sparse attention implementation that achieved 1.8x speedup over standard flash-attention for 32k context lengths.

Quantized Reasoning Model (QRM)

Role: Lead Researcher

Tools: JAX, Hugging Face, Docker, AWS

Designed a 4-bit post-training quantization pipeline for transformer models, preserving 98% of baseline accuracy while cutting host costs by 50%.

• Google Cloud Professional Machine Learning Engineer (2023)
• NVIDIA DLI Certificate – Troubleshooting Distributed Training (2022)

• Publications: 4 peer-reviewed papers in NeurIPS (2), ICML (1), and ICLR (1) as primary or secondary author
• Open Source: Contributor to Hugging Face transformers and Ray libraries, accumulating over 450 stars on public repositories
• Availability: 3 weeks notice (relocation to Bay Area preferred)