Projects

A cycle-accurate CPU and accelerator simulator for GPT-style forward passes, including systolic-array matrix multiplication, attention execution, SRAM/HBM transfers, and configurable tiling. Roofline and trace-driven analysis exposed the shift from compute-bound to memory-bound execution beyond 2K-token sequences.

?At what context length does moving transformer state cost more than computing on it?

HBM traffic at 8K

3.2× reduction

Throughput at 8K

1.8×

A100 estimate error

≤12%

An upstream MLIR lowering path for BERT-class transformer execution on AMD ROCm through IREE. The work fused LayerNorm–Dropout–Residual Add sequences, added cross-backend numerical validation, and extended end-to-end CI while collaborating with four maintainers across 12 pull requests.

?How much transformer overhead can disappear when the compiler sees across operator boundaries?

Kernel-launch overhead

~40% lower

Upstream collaboration

12 PRs

C++ + MLIR

3,800+ LOC

A Go CLI that injects controlled failures into Kubernetes and OpenShift test environments: pod termination, API latency, reconciliation delay, and simulated network disruption. Recovery is measured from controller logs, cluster events, resource state transitions, and Prometheus metrics, so failure modes in provisioning, deletion, and cleanup show up as data rather than anecdotes.

?Can a failure harness turn recovery behavior into evidence instead of anecdotes?

An operator that detects divergence between declared multicluster state and what is actually running, then reconciles namespaces, labels, policies, and workload configuration back into line. Built on Kubebuilder with finalizer handling, owner-reference validation, status conditions, structured logging, retry-safe remediation, and integration tests across local clusters.

?What does safe reconciliation require when declared and observed multicluster state diverge?

A reverse proxy with request routing, connection reuse, retries, timeouts, health checks, structured access logs, and upstream failure handling. Token-bucket rate limiting per route, per-route metrics, and throughput and tail-latency benchmarks run under concurrent load with wrk and vegeta.

?Which failure behavior belongs in the proxy instead of every caller?