Projects

Agentic Distributed System Ops

Order Lab @ UMich, May 2025–Present, Advised by Ryan Huang

agentic-ds-ops

Built an agent-based auto-mitigation loop (reproduce $\rightarrow$ measure $\rightarrow$ decide $\rightarrow$ mitigate) for common failures: overload, network latency/loss/partitions.
Targets: Apache ZooKeeper/HDFS (3-node dev clusters); faults via ChaosBlade/Chaos Mesh and tc network controls; synthetic workload generators.
Mitigation: HAProxy (TCP routing/failover) and Resilience4j (circuit breaker, bulkhead, rate limiter) with self-implemented scripted rollback/timeout policies.
Observability: Prometheus + JMX Exporter + ps/JVM Flags tracking SLO metrics p99_latency, outstanding_requests, and System resources metrics IFACE_throughput, cpu_usage

UMich CSE 582 Term Project (Advanced Operating Systems by Prof. Ryan Huang), Aug-Dec 2025

cuda-proxy-player

Designed a hybrid CUDA runtime combining CUDA Graphs for stable compute paths and persistent kernels for irregular on-device tasks.
Reduced kernel launch overhead and tail latency without brittle fusion, enabling dynamic routing and static graph reuse.
Implemented bucket-based static pooling, piecewise graph capture, and stream-level synchronization for efficient MoE execution.

UMich CSE 589 Term Project (Advanced Networking by Prof. Z. Morley Mao), Aug-Dec 2025

fastEMMA

Designed a hybrid local–cloud pipeline combining OpenEMMA and large VLMs for efficient & reliable decision-making.
Selective offloading of static-scene context (e.g., construction, obstacles) to cloud models; dynamic agents handled locally.
Network-aware control adapts send rate and compression based on latency to balance accuracy and responsiveness.

UMich EECS 482 (Introduction to Operating System with Special Topics), Jan–Apr 2025

Threads: user-level thread library (swapcontext/makecontext), Mesa monitors; interrupts/core-suspend; non-preemptive FIFO run queues. Thread library supports multiprocessor systems.
Pager/MMU: SWAP & FILE-backed pages; page tables with present/reference/dirty bits; COW lazy copying + Clock eviction; Supports fork/mmap for non-empty arena (process).
Network File System: inode-based, Unix-style NFS; Concurrency with Boost shared/unique locks and hand-to-hand locking mechanism;

EECS 489 (Introduction to Networks), Jan–Apr 2025

Mininet-style iperf: Custom Mininet topologies(python) with iperf-style RTT/throughput measurement(C++ sockets);
- Reproduce the buffer-bloat problems with long switch queue and asymmetric link capacities. Solved with ovs SDN control with 2 prioritized queues. (This belongs to CSE 589 project)
Video Server emulation: Reverse proxy via select+poll + round-robin load balancing; DASH adaptive streaming.
Transport Layer emulation: TCP sliding window over UDP-style protocol (implemented GBN/SR rules, cumulative/selective ACKs);
L3 router simulation with ARP, ICMP, basic intra-domain routing.

EECS 388 (Introduction to Computer Security), Jan–Apr 2025

Digital Forensics: disk image analysis, file carving, metadata extraction with Steganographic Tools;
Cryptanalysis & cracking: length-extension, padding-oracle; John the Ripper (break encrypted PDF/ODT), Hydra (SSH).
Web exploitation: auth bypass via XSS/SQLi/CSRF, write a simple “Samy-is-My-Hero” Worm Attack.
Binary Attack (stack overflow attack): x86 assembly analysis and exploit development workflows using gdb;
- focused on stack-overflow patterns, return-oriented techniques (ROP chains), and bypass mitigations like ASLR and Stack Canary.
Reverse Engineering: recover executable → C source logic (decompilation & manual analysis) with Ghidra and static analysis;
- combined with pwning to hack the UMich BBB Building’s stair light

HDR Lab @ UMich, Sept 2024–Present, Advised by Xiaonan Huang

origami-robotic-arm

Kresling origami + pneumatic actuation for confined-space manipulation.
Led STM32 & Orange Pi firmware (dynamics/PID, CAN/I²C); C++/Rust core algorithms; contributed to PCB design.
Demos/extended abstracts: ICRA 2025 Workshop (Best Poster), RoboSoft 2025 Workshop, ICON 2025.

Sirius Lab @ SJTU, Feb–Sept 2024, Advised by Yutong Ban

Built an LLM-driven pipeline that turns natural-language instructions as well as RGBD visual inputs into executable manipulation plans on a Flexiv 7-DoF manipulator.
Implemented FK/IK and motion planning with Flexiv RDK; integrated ZED depth perception for scene understanding and grasp placement.
Designed a constraints-aware controller (joint/velocity/workspace limits) with smooth trajectories.
Combined simulation data with real runs to improve task reliability.

RoboMaster @ SJTU, 2023–2024, Won National Championship in the 2023 RoboMaster Competition (Awarded by DJI)

sentry

STM32-F407 platform; dual-gimbal stabilization with optimized PID controller; high-response 4 steering wheel chassis.
CAN/UART pipelines for CV/LiDAR;
IMU-based absolute-pose control.
High precision shooting with flywheel speed & lead-angle compensation.