const arvin = {
name: "Arvin Wang",
role: "Infrastructure Engineer",
direction: "AI Infrastructure",
focus: ["Reliable systems", "Automation", "Observability"],
building: ["ChatPassport", "nccl-allreduce-bench", "softroce-rdma-nccl-lab"],
};- π Hi, I'm Arvin. I am working toward building reliable, observable, and automated infrastructure for AI systems.
- βοΈ My engineering interests span network reliability, infrastructure automation, observability, and distributed systems.
- π§ I am especially interested in AI infrastructure, model training platforms, and reliability engineering.
- π οΈ This profile is a public workbench for open-source projects and engineering experiments.
- AI Infrastructure β training systems, resource efficiency, and scalable execution.
- Reliable Systems β high availability, observability, capacity planning, and incident response.
- Infrastructure Automation β turning operational workflows into repeatable tools and pipelines.
- Open-source Building β creating practical tools around AI workflows and user-controlled data.
- SQLFluff β PostgreSQL 18 temporal foreign key parsing. #9150 Β· Merged Nov 14, 2026
- mycli β Fix DSN encoding for usernames containing
/, with regression tests. #2446 Β· Merged Nov 18, 2026
Carry the context. Keep control.
Continue AI conversations across ChatGPT, Claude, Gemini, and DeepSeek while keeping the context local and the user in control.
- Local-first, with no developer-operated server, account, analytics, or tracking.
- Supports structured conversation import, export, and cross-platform continuation.
- Built with TypeScript, React, WXT, Zod, Vitest, and Chrome Manifest V3.
Understand the curve, not just the number.
Controlled NCCL allreduce experiments on two Tesla T4 GPUs show how topology, transport, algorithm, and protocol shape the bandwidth curve.
- Changes one NCCL setting at a time and checks diagnostic logs to confirm it took effect.
- Includes timing logs, analysis scripts, and plots so the results can be reproduced and inspected.
- Examines SHM fallback when P2P is unavailable, plus Ring and Tree performance on the tested virtualized platform.
Follow the packets, verify the transport.
Two cloud GPU VMs running software RoCEv2, with evidence from verbs connectivity through NCCL AllReduce.
- Confirms RoCEv2 traffic with a 200-packet capture and NCCL data channels using
NET/IBoverrxe0. - Compares RDMA and TCP Socket AllReduce on two RTX A4000 GPUs, with zero validation errors in both recorded scans.
- Publishes 44 original evidence files, derived tables and plots, reconstructed run scripts, and documented failures and measurement limits.

