Gary Geng

LLM inference kernel optimization on Hopper & Blackwell, using CUTLASS, Triton, CUDA/PTX, JAX. OAI Triton compiler development & optimization. Low-precision fused MoE optimizations.

Due to non-compete agreement

Worked in the AWS Chime Telephony team to implement advanced metric analysis for the SIP trunking service. Used AWS CloudWatch Contributor Insights to categorize service failures by customer, error type, and server instance, creating specialized metrics that are more accurate and reliable. Using the new metrics, created more accurate alarms that 1) reduced false positives by ~80% and 2) enabled quicker detection of failure for smaller customers, reducing failure response time by typically 30-95%. After surveying senior teammates for common error scenarios, created an extensible dashboard using aggregate metrics that enables on-call engineers to quickly detect the scope, nature, and cause of failures.

Developed Stibium, a parse, autocompletion, and static analysis tool for Antimony: https://github.com/evilnose/stibium Developed BioIDEK, a VSCode language support extension for Antimony: https://github.com/evilnose/vscode-antimony Wrote custom error recovery logic for the parser that backtracks the parser’s history states, so that the entire source document can be analyzed at once, despite user syntax errors.

Developed a robot simulation engine for testing path-finding algorithms. Implemented two pathing algorithms with performance-optimality trade-offs for the competition. Used C++ profiling tools to improve map partitioning performance by 4x and A* pathfinding performance by 3.5x.

Coordinated the autonomous, networking, object detection, and UI sub-teams (~20 people in total). Hosted workshops on C++ and web frontend (React) for new members.

Was part of an ongoing weather data visualization project and worked on several tasks, including: improving and implementing features in the web frontend (Angular), adding HTTP endpoints for database querying (MySQL), and processing and visualizing raw weather data in Javascript and Python. Also implemented a clustering algorithm that groups days with similar weather characteristics out of more than 1000 days (Python).

Worked with Dr. Hesam Jahanian and two others on implementing a convolutional neural network (U-Net) for synthesizing brain PET scan images from MRI image inputs. Was responsible for preprocessing and visualizing the image data for training, setting up the training environment using Docker, and tuning the neural network hyperparameters.