Review and optimize kernels and data movement for CUDA, Triton, Metal, OpenCL, SIMD, and accelerator-style workloads. Use when diagnosing throughput gaps, memory bandwidth limits, launch overhead, occupancy issues, kernel fusion tradeoffs, or host-device transfer bottlenecks in ML and systems research.
Build evaluation plans and error-analysis workflows for ML, retrieval, generation, systems benchmarks, and embedded or perception pipelines. Use when adding metrics, checking regressions, designing ablations, interpreting leaderboard changes, or debugging why a model improved on one slice and failed on another.
Rapidly map an unfamiliar research or engineering codebase, identify entry points, execution paths, configuration layers, tests, and risky modules. Use when inheriting a project, preparing a reproduction, reviewing a repo for collaboration, or locating where to modify a model, runtime, compiler, or hardware flow.
Analyze compiler, runtime, and code generation systems including IR lowering, scheduling, memory layout, graph compilation, autotuning, and runtime overhead. Use when profiling a compiler stack, comparing generated code, debugging performance cliffs, or evaluating compiler research claims.
Design and debug state estimation, filtering, system identification, and control-oriented models for robotics, autonomous systems, and embedded control. Use when choosing observers, Kalman variants, sensor fusion structure, stability assumptions, or diagnosing drift, lag, and closed-loop estimation failures.
Create and audit datasets for CS and EE research, including schema design, collection pipelines, deduplication, split strategy, leakage prevention, labeling QA, licensing, and provenance tracking. Use when building a dataset, merging corpora, preparing train, validation, and test splits, or validating a benchmark before publication.
Debug distributed systems behavior including consistency issues, queue backlogs, retries, partitions, replica divergence, tail latency, and backpressure. Use when a service mesh, stream processor, storage system, scheduler, or multi-node research system behaves differently under scale than in local tests.
Design and review digital signal processing pipelines including sampling, filtering, transforms, detection, estimation, feature extraction, fixed-point concerns, and implementation tradeoffs. Use when developing or debugging DSP methods for communications, sensing, audio, imaging, robotics, or embedded systems.