Hi, I’m Nancy. I build and debug AI systems from the infrastructure up - from CUDA and GPU environments to inference runtimes and agentic workflows. Most of my projects begin as experiments: I build, break, test, and turn what I learn into repeatable workflows. I bring 10+ years of earlier experience in enterprise infrastructure, data systems, high-availability operations, and incident response.
- vLLM — PR #34978: Investigated a DeepSeek-R1 reasoning-boundary failure involving speculative decoding and structured output, then submitted scheduler/parser changes and validation for the failure path.
-
GPU platform reliability: NVIDIA and AMD/ROCm environment qualification, hardware/runtime compatibility, Kubernetes GPU scheduling, observability, and failure recovery.
-
Inference systems: vLLM and Triton serving, workload characterization, TTFT/ITL/throughput analysis, KV-cache behavior, routing, fallback, and backend health.
-
Reliable AI workflows: evidence contracts, bounded agent execution, evaluation, provenance, and human approval boundaries.
-
Frontier systems exploration: world-model training and inference infrastructure, accelerators, reliability, and evaluation.
📖 I write about what I’m building and learning at StackBytes.
📬 Contact: nfrounjian@gmail.com | Connect on LinkedIn