Trinadh Musunuri

IBM — Austin, TX

• Engineered a multi-agent AI platform using React and IBM watsonx Orchestrate, integrating Slack delivery with structured interactive widgets for enterprise workflows.
• Built transaction-centric Grafana and Loki dashboards for distributed voice workflows, enabling end-to-end tracing across AI Gateway, STT, TTS, and downstream services — reducing root-cause analysis time by approximately 80% during internal investigations.
• Implemented structured logging and contextual metadata propagation (transaction ID, request type, provider, latency metrics) in AI Gateway, improving request correlation and operational visibility across services.
• Introduced service-level failure drill-down and STT lifecycle event classification, simplifying debugging across multi-service AI voice workflows.

Central Michigan University

• Replaced an evolutionary algorithm with a gradient-based optimization method, achieving a 50× faster runtime (30–60 min → under 3 min) on adversarial ML experiments.
• Conducted low-level architectural analysis on AMD Ryzen AI hardware (XDNA 2 / Krackan), applying fine-grained cycle counting and DRAM contention metrics to profile NPU execution patterns.
• Designed and executed CPU–NPU benchmarking scripts to collect high-fidelity performance data across deep learning workloads, contributing directly to hardware security research publications.
• Engineered an interactive adversarial YOLOv2 attack demo — currently in active use by students — hosted on Hugging Face Spaces with a Cloudflare Worker proxy to work within restrictive university firewalls.