Ray Shao Yingzhan
Profile
Cloud Infrastructure & Platform Engineering student building and operating reliable production systems at the intersection of cloud computing, AI/ML workloads, and developer tooling. Incoming Site Reliability Engineer Intern on Tencent’s AI Compute Centre, operating AI-accelerated HPC clusters (InfiniBand/RoCE, Slurm, Kubernetes). Manages GCP infrastructure at Google through incident response and root cause analysis, built CI/CD pipelines and internal platforms at AMD, and operates self-hosted Kubernetes clusters. Targeting Cloud Engineering, SRE, DevOps, and Platform Engineering.
Education
Nanyang Technological University, Singapore | Aug 2023 to Aug 2027
Bachelor’s in Computer Engineering (Honours)
Final Year, Graduation Aug 2027 · Available for internships from Jan 2027 · Available for full-time roles from May 2027
Modules: Distributed Systems, Operating Systems, Computer Networks, Cloud Computing & Big Data Systems, Embedded AI Systems, GPU Programming, Machine Learning, Microprocessor System Design & Development
Final Year Project (Industry-Sponsored, AMD)
Experience
Tencent
Site Reliability Engineer Intern, AI Infrastructure (Incoming), AI Compute Centre, Overseas IT | Jan 2027 to May 2027
- AI HPC Operations - Joining the AI Compute Centre to support daily operations and maintenance of AI-accelerated high-performance computing (HPC) clusters across global regions
- Cluster Delivery - Supporting deployment, configuration, and maintenance of high-end compute, storage, and networking across bare-metal, HPC, Kubernetes, and Slurm environments
- High-Speed Networking - Checking cluster network connectivity across InfiniBand and RoCE links, running predefined tests, and flagging link and performance anomalies
- Observability - Monitoring compute, storage, and management network dashboards (Prometheus, Grafana, ELK), recording bandwidth, latency, and RDMA results against baselines
- Incident Support - Supporting incident investigations by collecting switch, NIC, and host logs, updating incident records, and tracking vendor follow-up actions
Xiaomi
Big Data Site Reliability Engineer Intern | Sep 2026 to Dec 2026
- Distributed Systems Reliability - Supporting reliability of production big-data infrastructure (HDFS, HBase, ElasticSearch, Doris, RocksDB, ClickHouse) underpinning Xiaomi’s AIoT platform at scale
- Automation Tooling - Building efficiency tooling to automate management and operations of large-scale distributed systems, reducing manual intervention on recurring tasks
- On-Call Operations - Participating in 24x7 on-call rotation, triaging and resolving production performance degradations and system failures
- Agentic On-Call Tooling - Developing a Python MCP server that lets an LLM agent triage on-call incidents: it searches Grafana Loki logs, runs host diagnostics (strace, perf, tcpdump, systemd), routes alerts to the right runbook, and triggers routine operations through Ansible
- Cross-Team Reliability - Partnering with application development teams to troubleshoot cross-system issues and drive infrastructure reliability improvements
Data Center Engineer Intern, Global Server Operations, Google Cloud | May 2026 to Aug 2026
- Observability - Built an end-to-end fleet monitoring pipeline (BigQuery SQL transforms to Looker Studio serving layer) surfacing cross-site performance and time-series trends across multiple datacenter sites; cut MTTR by 40% and was adopted by senior leadership as the at-a-glance operational review surface
- AI Tooling - Developed an in-house AI diagnostic agent automating fault triage across production infrastructure, cutting manual investigation steps for recurring failure patterns
- Cloud Reliability - Drove postmortem-based root cause analysis on production infrastructure faults, isolating fault domains and eliminating systemic failure patterns across compute, storage, and networking
- Cloud Networking - Operated production network fabric across Cisco, Arista, and Juniper, diagnosing 100G-800G interconnects, QSFP optics, and fiber plant to restore link availability against SLO targets
- Platform Onboarding - Drove qualification and validation of new hardware and network platforms toward sustained mass production, coordinating across engineering teams and global partners
- Incident Response - Restored connectivity for 50+ incidents across the network fabric and reduced the repeat incident rate by 10% through postmortem-driven remediation
- Incident Management - Prioritized maintenance by workload criticality and error budget impact, reducing cascading failures across production services
- Community - Designed and led a Google Developer Group Singapore workshop on Hermes (Nous Research autonomous agent framework) on 23 Sep 2026 for ~175 attendees (~200 registered)
Advanced Micro Devices
Product Development Software Engineer Intern, Data Center Group (DCG) | May 2025 to Jan 2026
- Data Platform & Tracing - Cut data-processing time from 70 to 5 minutes by replacing a failing legacy ETL pipeline (2-3 missed runs weekly) with a modular Extract/Transform/Load platform (Python + SQL); implemented granular execution tracing across batch stages and real-time observability dashboards, reducing debug cycle time by 60% for 10+ engineers
- Discovery to Delivery - Found workflow pain points across the HTOL team beyond the assigned project and shipped fixes: a lot-split scanner (70% less time, 100% adoption in 2 weeks), a Kanban lot tracker (60% fewer tracking inquiries), and the text-to-SQL chatbot
- Customer Engagement - Represented AMD to prospective customers and users at SWITCH 2025 for 3 days; organized a cross-site session so other divisions could adopt an internal agentic AI tool
- Platform Engineering - Built internal platform service (FastAPI + text-to-SQL) enabling self-serve infrastructure telemetry from distributed data lake, driving 3x increase in platform adoption
- CI/CD - Built automated data processing pipelines (Python, custom tooling) for cloud hardware validation, reducing release cycle time by 45% and enabling continuous delivery of qualified cloud platforms
- Enterprise Agentic Platform - Architected a multi-persona digital twin agentic platform (FastAPI + Weaviate hybrid RAG + RBAC) using custom state machines and Pydantic schema adapters; implemented cyclical decision loops, tool-calling handoffs, and sandboxed SRE automation across 15+ engineers with zero hallucination escapes
- Quality Engineering - Ensured cloud GPU platform reliability for MI350/MI400 datacenter accelerators through structured failure analysis and SLO-driven pre-production qualification, achieving zero escapes to deployment
- Full-Stack - Built internal platform (React + FastAPI) with RBAC and performance monitoring, deployed across multiple engineering teams for event management and cross-team collaboration
Projects
High-Concurrency LLM Serving on AMD MI50 (vLLM, ROCm) | Jul 2026 to Present
- Inference Throughput - Ported vLLM continuous batching and PagedAttention to AMD Instinct MI50 (gfx906), reaching 857.5 tok/s aggregate at 24 concurrent streams
- Tensor Parallelism - Built a heterogeneous TP=2 engine (MI50 + RX 6900 XT) with a patched Triton attention kernel and multi-target RCCL: 189 tok/s on Qwen3-8B at 16 clients, 94% linear scaling
- GPU Isolation - Isolated GPUs in unprivileged Proxmox LXC containers using cgroup v2 device passthrough (/dev/kfd); monitored serving SLIs with Prometheus and Grafana
Sim-to-Real Autonomous Vehicle Stack (ROS 2, Zenoh, Gazebo, STM32) | Aug 2026 to Oct 2026
- Architecture & Integration - Architected and integrated the full ROS 2 Jazzy stack (7 packages) across a Raspberry Pi and a laptop over rmw_zenoh, with YOLOv8 perception on NCNN, and debugged it on hardware; ~78% of the team’s commits (378 of 486)
- Control - Replaced the reference move-and-stop firmware with 40 Hz velocity control on FreeRTOS and a pure-pursuit tracker, removing a 0.64 to 0.96 s stop after every motion primitive
- Telemetry, Planning & Testing - Built the Foxglove telemetry layer and rosbag recording (~125 runs); exact TSP with Reeds-Shepp and Hybrid A* paths; scored 36 headless Gazebo scenarios against ground truth using the real bridge and controller; ~332 tests
Distributed Big Data Engines (Hadoop, Spark, Kubernetes) | Jan 2026 to May 2026
- Hadoop MapReduce patterns in Java, PySpark jobs, AWS EC2 scaling benchmarks (Sysbench, iPerf3), and Kubernetes scaling (3.8x on 4 replicas, <2.4 s pod self-healing)
AMD FYP | Aug 2026 to May 2027
Industry-Sponsored Final Year Project, NTU + AMD
- Developing a computer vision anomaly detection pipeline for semiconductor defect classification, building on prior AMD internship domain knowledge in wafer test data and failure analysis
Self-Hosted Cloud Platform: Proxmox + Kubernetes | Mar 2025 to Present
- Infrastructure Automation - Reduced operational toil by 40% across self-hosted cloud by automating backups, log rotation, service orchestration, and runbooks on a self-hosted AMD EPYC Proxmox platform
- IAAS - Deployed full-stack observability platform (Telegraf/InfluxDB/Grafana) across 20+ containers and VMs, providing real-time SLI metrics for CPU, GPU, memory, and network utilization
- Container Networking - Architected multi-VLAN network topology and implemented container networking (Docker, LXC) across Linux/Windows VMs, enabling secure multi-tenant workloads and network isolation
VLASH-Forge / VLASH-Piper | Jan 2026 to May 2026
- Distributed Systems (VLASH-Forge) - Built portable training multi-node framework deployable across SLURM HPC, Docker Compose, and Kubernetes, with Hugging Face Hub versioning for reproducible ML experiments
- GPU Optimization (VLASH-Forge) - Reduced VLA fine-tuning VRAM floor to 12 GB (LoRA) / 8 GB (QLoRA) on consumer GPUs from multi-GPU full-precision baseline by sharing visual encoder weights across adapter steps
- Edge AI (VLASH-Piper) - Achieved the first known π₀.₅ deployment on Jetson AGX Orin by porting Temporal Delay Augmentation (VLASH, MIT Han Lab) and resolving JetPack/CUDA aarch64 build incompatibilities, cutting inference latency 29.5x (184.5ms vs 5,444ms) to sustain 30 Hz real-time robot control
- Robotics Integration (VLASH-Piper) - Lifted pick-and-place task success from 5% to 65% on embedded hardware; full pipeline from kinesthetic demonstrations to distributed training to edge deployment with edge inference at runtime
DA3 Underwater: LoRA Depth Estimation for AUV | Jan 2026 to May 2026
- Computer Vision - Fine-tuned Depth Anything 3 via rank-8 LoRA targeting all DINOv2-L attention projections (9.24% trainable parameters) on a single A100, reducing underwater monocular depth AbsRel from 0.774 to 0.320, a 59% improvement on a held-out cross-environment benchmark the model had never seen during training
- Preprocessing Pipeline - Applied physics-aware preprocessing (Gray World white balance, percentile histogram stretching) as a fixed input contract at both train and inference time to compensate for wavelength-dependent light attenuation rather than leaving it to the model
- Edge Deployment - Merged LoRA weights into base model for TensorRT compilation, removing adapter scaffolding from deployment artifact; integrated with YOLOv11-seg into ROS 2 Jazzy AUV pipeline for SAUVC and RoboSub
Mixed-Precision MAC Unit: TSMC 180nm ASIC | Jan 2026 to May 2026
- Implemented a merged FP16×FP16→FP32 / dual-INT8 MAC datapath across 15 Verilog-2001 modules and took it through full RTL-to-GDSII on TSMC 180nm using Synopsys Design Compiler and Cadence Innovus, closing post-layout timing at 157 MHz
- Identified an undocumented bit-width assumption in the IEEE ISCAS 2018 Karatsuba formulation by running 100,000 test vectors against a bit-exact Python reference model with 24 parallel Icarus workers, correcting a 47% floating-point failure rate that the paper’s original STM-90nm test suite never exposed
- Achieved 4.6% area overhead versus a standalone FP16 MAC while adding a parallel dual-INT8 path, 21% smaller than a naive side-by-side implementation by sharing Karatsuba sub-multipliers across both precision modes
Leadership & Co-Curricular Activities
NTU Semiconductor & Compute Club | Aug 2025 to Present
Founder and President
- Spotted a gap between NTU’s curriculum and AI-economy roles across the compute stack (SRE, inference, GPU, chip design, fab) and founded the university’s first club covering it; reached 100 members in one week and won official recognition in Jul 2026 after building traction without approval
- Grew to an 800+ member community (300+ registered); drove scrappy 0-to-1 execution to secure partner-hosted and sponsored technical programs with 7+ organisations including AI Singapore, AMD, A*STAR IME, GlobalFoundries, Micron, SSIA and ASMPT
- Built the organisation from scratch: Exco plus four portfolio teams (Events, Biz Dev, Publicity, AI Product & Platform), a 29-person committee, and a panel-interview recruitment process with a shared scorecard (10 openings in the AY26/27 intake)
- Lead a 6-person AI Product & Platform team (2 each of AI SRE, AI FDE, AI PM), placed at the centre of club problems to ship tools: a Notion admin agent (estimated ~4 hours of committee admin saved per week), a daily AI news briefing agent for an 800-member Telegram channel, and the club website (built myself)
- Set standards through expectations at hiring, performance reviews and coaching; after a partner event fell short of its target, assigned clear owners for external communication and introduced attendance deposits
Singapore Semiconductor Industry Association (SSIA) | Jan 2025 to Jan 2026
SAY Ambassador
- Championed SSIA flagship events and the IC Design Summer Camp, working with senior industry executives to build technical mentorship pipelines
Technical Skills
Cloud & Infrastructure: GCP, Linux (strace, perf, tcpdump, systemd, cgroups v2), Ansible, Proxmox, Docker, Kubernetes, KVM/QEMU/LXC, VPC/VLAN networking, load balancing, high availability (RAID, failover, disaster recovery)
DevOps & SRE: GitHub Actions, Prometheus/Grafana (Telegraf/InfluxDB), incident response, postmortems, SLO/SLA monitoring, error budgets, capacity planning, runbook automation, Terraform (learning, in progress)
Software & Data Engineering: Python, SQL, C++, TypeScript, FastAPI, React, ETL, Pandas, NumPy, Git
AI & ML Systems: vLLM, PagedAttention, ROCm, RCCL, Triton, ROS 2, PyTorch, TensorFlow, Weaviate hybrid RAG, distributed training (DeepSpeed, FSDP), TensorRT, ONNX, edge inference (Jetson AGX Orin)
Certifications
Google Professional Certificates: IT Support, IT Automation with Python, Cybersecurity, Advanced Data Analytics
WSQ Employability Skills: Personal Effectiveness, Organizational Relations, Visionary Leadership
Awards & Achievements
- Edusave Character Award (2020): Recognized for leadership and contributions to community
- Merit Prizes: International Biomedical Quiz (2020), Science Chronicles Writing Competition (2020)
- Best Cinematography: National Chinese Short Film Competition (2018)