• Characterized the Llama-2-7B inference workload for the Mugi LLM accelerator by classifying all 2,500+ per-pass operations into four hardware execution buckets, then generating a validated ONNX graph consumed by the lab’s cost/cycle simulator.
• Rearchitected the Python/PyTorch profiling pipeline to replace per-model if/else dispatch with a plugin adapter registry, uncovering and fixing a latent binning-format defect that had silently broken instrumentation for every model outside the one hardcoded configuration, unblocking profiling for additional architectures.
• Designed a one-shot profiling API producing per-layer lookup-table configs and hardware workload specs from a single config object, through a pluggable strategy/criterion sizer validated end-to-end against the simulator on an NCSA Delta HPC cluster, collapsing per-layer hand-editing into one command.
• Designed an in process evaluation harness that repatches lookup tables in place rather than reloading a 13GB model per layer, converting a 32 round manual search into one unattended job that sweeps a 16-window grid at every layer, seeded per layer from measured exponent histograms.