Open to opportunities

Punyaa Minocha

@punyaaminocha

Message

AI systems and kernel engineer accelerating quantized deep learning on edge hardware.

India

Message

What I'm looking for

I’m looking to work on hardware-software co-design for edge AI—building quantized kernels, speeding up transformer/LLM inference, and validating performance from cycle-cost analysis through end-to-end releases.

I build and optimize high-performance quantized deep learning kernels for custom edge AI accelerators at Kinara AI (acquired by NXP Semiconductors), using C++ and Python. I’ve implemented operators like Softmax, Convolution/Depthwise Convolution, MatMul, Batch Normalization, and elementwise ops, and used instruction-level tuning and operator fusion to speed up production inference workloads.

On transformer and LLM deployments for resource-constrained environments, I’ve driven INT4/INT3/INT2/INT1 quantization, KV-cache optimization, speculative decoding, and dynamic kernel implementations. I also supported ASIC bring-up, silicon correlation, end-to-end network validation, and unified kernel build/validation workflows, including cycle-cost analysis and instruction-slot tuning for LLaMA2 and LLaMA3.1.

Experience

Work history, roles, and key accomplishments

Current

Software Developer (G1)

Current

Kinara AI

Aug 2023 - Present (3 years)

Develop and optimize high-performance quantized deep learning kernels for custom edge AI accelerators using C++ and Python. Achieved over 50% improvement in inference throughput across production workloads.