Qualcomm — Rohin Sood

Hardware investigations leave a lot of evidence behind: logs, subsystem behavior, and the context engineers bring to a failure. My work at Qualcomm connected that engineering context to agentic software—tools that could help people investigate and get work done.

A platform for repeated work

I created a platform for building and deploying agentic workflows, supporting more than 15 deployed modem software-engineering jobs for over 600 users. The backend powered natural-language builder chat, agent creation, and LLM-as-judge evaluation. The useful unit was a workflow an engineer could actually use, rather than an isolated model response.

Keeping the hardware context

I also optimized agent harnesses for root-cause analysis across modem, camera, CPU, and GPU logs, and contributed to LogTalk’s post-training process. This connected the software around a model to the engineering evidence it needed to investigate.

The device side of AI

My résumé also covers Snapdragon X Elite NPU benchmarking for MCP and RAG agents, including C/C++ QNN and Hexagon API integration for tensor allocation, operator dispatch, and device synchronization. The reported benchmark improvements were 3× decoding throughput and 34% lower time to first token. Those are results from that benchmark work, not a claim about every model or workload.

The connection I care about

This is where my interests in enterprise software and intelligent devices meet: the agent needs useful tools and evaluation, while the underlying compute still determines how quickly it can respond. I want to keep working across that boundary.