Muneeb Ahmad

San Diego, CA

• Analyzed ray-tracing performance on Adreno GPUs for both Vulkan and DirectX12, on both ray pipeline and ray query workloads.
  • Identified bottlenecks using hardware performance counters and traced fixes back to the shader compiler codegen.
  • Fixes resulted in a 2x speedup in targeted configurations.
  • Identified memory overallocation bugs in the compiler.
• Proposed memory layouts that achieve 100% (2x existing) memory bandwidth utilization for ray-tracing internal data by improving coalescing behavior; the design also improves cache-friendliness.
• Created AI agent-based tooling to aid in end-to-end on-device performance analysis, regression testing, and root-causing.

UC Davis

• Created optimized custom GPU kernels for compressed paged attention with fused RoPE using DSLs at various levels of abstraction Triton, Gluon and HipKittens.
• Integrated custom attention paths into vLLM, including managing the KV cache in bespoke formats.
• Implemented gesvda on top of rocSOLVER for 10x faster approximate singular value decomposition on AMD GPUs.
• Developement of hardware-efficient methods to compress the KV cache 8–30x with minimal accuracy degradation (<2% on task performance).
• Extensively evaluated LLM task performance using models like Llama and Qwen3 on benchmarks like LongBench and GSM-8K.

University of New Mexico (Remote)

• Created a framework to find the optimal deployment strategy for a given deep learning model on a given edge device, targeting satellites and robotics.
• Achieved a ~400% improvement in latency on the Coral Edge TPU dev board.
• Worked with hardware accelerators such as the Coral Edge TPU and NVIDIA Jetson Nano.
• Applied quantization to models and transformed formats for compatibility with different hardware.

University of California, Davis

• Advisor: John Owens
• Coursework: Modelling & Optimization in Computer Engineering, Large-scale Scientific Computing, High Performance Statistical Computing, Computer Architecture, Information Theory & Coding

Jamia Millia Islamia, New Delhi, India

• CGPA: 9.05 / 10
• Coursework: Computer Networks, Network Security, Computer Architecture, Digital System Design, Embedded Systems, Operating Systems, Parallel & Distributed Computing