I started as a full-time intern in 2016 and have worked in software engineering since graduating in 2017. My work spans complex backend systems, distributed systems, and high-performance engineering, and I now work in quantitative software development.
My primary stack is C# and .NET, with additional production work in C++, Rust, and compute infrastructure.
My work on backtesting and live trading systems spans C# / .NET, C++, and Rust, including matching, deterministic parallel replay, model inference, and performance optimization, with output parity checks used to validate correctness.
Market Data Engineering
I build market data pipelines for transcoding, columnar storage, streaming reads, incremental scheduling, and cross-data-center synchronization, and maintain a ClickHouse ingestion pipeline for production trading logs.
Backend & Distributed Systems
C#, .NET, and ASP.NET Core are my primary stack. My production work spans high-throughput, high-performance backend and distributed systems, including caching, messaging, RPC, distributed consistency, high availability, and disaster recovery.
Compute Infrastructure & Reliability
I contribute to cross-data-center compute infrastructure spanning Slurm scheduling, GPU compute, parallel filesystems, identity services, node automation, monitoring, and incident response across compute, storage, and networking.
Correctness & Engineering Governance
I use output parity checks, automated testing, code review, and independent audits as code-acceptance gates. I maintain AI-assisted development workflows through structured handoffs and shared tooling configuration and upgrade scripts, while contributing to technology selection and engineering standards.
Experience
A Quantitative Investment Firm
Quantitative Developer
Develop and maintain backtesting and live trading systems, market data platforms, and compute cluster infrastructure, using C# / .NET, C++, and Rust to support research and production workloads.
9.37 secondscached full-market replay for one trading day across 4,932 stocks
7.7×faster than the C# backtesting baseline
2.4–2.8×faster reads for monthly market data workloads
Byte-identicalvalidated backtest and matching outputs
Backtesting & Live Trading Systems
Develop C# / .NET, C++, and Rust systems for backtesting and live trading. A new Rust engine replayed one trading day for the full market (4,932 stocks) in 26.85 seconds, or 9.37 seconds with cached inputs. On a separate 2,033-stock workload, the same task fell from a 28.9-second C# baseline to 3.77 seconds, a 7.7× speedup, with byte-identical outputs verified by SHA-256. This shortens full-market replay and result-validation cycles for research iteration.
Built a C++ batch matching engine processing about 2.4 million events per second for the Shanghai market. On 32 threads, it processed a full trading day for 2,306 Shanghai-listed stocks in about 18.9 seconds and produced 15.47 million trades byte-identical to the existing implementation.
Designed a streaming parallel replay pipeline with run-to-run byte-identical multithreaded results as a hard determinism constraint. A full day across 2,306 stocks, including inference, completed in about 52 minutes, or approximately 4.6× real-time speed.
Optimized the live market data path, reducing key-thread stalls within each model interval from 3,795 ms to 0–1 ms while preserving bit-identical output across 119,669 rows. Core-pinned single-thread throughput covered observed production peaks.
Completed a full-day replay at real-time speed (1×) for an extreme session covering 6,600 instruments and more than one billion events, with zero ring overruns. Reduced inference from about 4 ms in ONNX Runtime to 66 μs (269 μs for a monolithic TensorRT engine and 66 μs after parallel graph partitioning); end-to-end latency, including feature packing, fell from 327 μs to 124 μs, with a 65× improvement at p99.
Market Data Platform
Migrated full-history order, trade, and snapshot data for Shanghai and Shenzhen equities, together with futures data, from CSV gzip files to Parquet and Arrow IPC. Implemented C# and Rust transcoders, incremental scheduling, cross-format consistency checks, and malformed-data recovery.
Optimized streaming market data loading for monthly datasets exceeding 10 billion rows, making reads 2.4–2.8× faster, reaching 1,779 MB/s peak throughput, and reducing peak memory use from about 40 GB to 2–4 GB. This reduces I/O and memory pressure for large-scale market data research workloads.
Process about 100,000 data objects in each daily increment. The post-market data batch covers 50,000–70,000 objects and completes in 2–3.5 minutes; the full workflow, including validation and cross-data-center synchronization, completes in about 20 minutes.
Maintain a scheduled pipeline that ingests production trading logs into ClickHouse, with result notifications, failure detection, and reruns. It scans every minute and looks back three days for late-arriving logs.
Compute Infrastructure & Engineering
Build and maintain cross-data-center compute infrastructure covering Slurm scheduling, hundred-GPU-scale resources, parallel file systems, and identity services; delivered dual-controller high availability, weighted CPU/GPU partition scheduling, and controller migration.
Automated and standardized node provisioning, retirement, and runtime configuration with Ansible and scripts. Added checks and alerts for node state, GPU health, OOM conditions, and filesystem capacity, and resolved compute, storage, and network incidents.
AI Engineering: Established governance and validation practices for AI-assisted development: cross-checked agent implementations, independent audits, and structured handoffs, with result comparison, automated tests, and independent review built into code acceptance; maintain shared scripts for team-wide AI tooling configuration and upgrades.
Primary technologies: C# / .NET, C++, Rust, Flink, ClickHouse, Arrow, Parquet, Slurm, BeeGFS, MinIO, Ansible, and CUDA / TensorRT.
Tongcheng Network Technology Co., Ltd.
Senior Engineer, Domestic Flight Search (T6)
Worked on Tongcheng's core domestic flight-search platform, supporting tens of millions of active users and multi-channel distribution traffic through more than 1,000 production service instances.
1,000+service instances
28,000+CPU cores
35,000+ QPSpeak request rate
1,000+ TPScore API throughput
~150 ms avg / ~500 ms P95query path latency
99.9%annual availability
Children's fare pricing: Owned and redesigned complex legacy fare logic as a pipeline, improving runtime performance, code reuse, and workflow flexibility.
Low-fare calendar: Led a refactor to support finer-grained fare data, increasing low-fare accuracy from about 93% to more than 98%.
Round-trip fare calculation: Served as the secondary owner and worked with the team to optimize the query flow, reducing average computation time from about 800 ms to about 400 ms.
Observability and resilience: Instrumented key metrics through APM, structured logging, dashboards, and alerts, and improved circuit-breaking and degradation strategies. The core system remained 100% available during a network-isolation drill; during a single-data-center drill, the remaining half of the fleet sustained the full production load.
Primary technologies: .NET 8, Redis, Docker, APM, and distributed service governance.
Price Monitoring and Alerting System
Dec 2023 — Feb 2025 · Primary owner of architecture and key modules
Validated real-time query prices against historical baselines and enforced safeguards across the full booking flow, from automatic filtering in search to blocking at order placement. Anomalies triggered phone or WeCom alerts within seconds.
1B+ per dayprice checks
~300K/secpeak validation rate
< 1%steady-state false-positive rate
100%anomalous-product blocking
Owned the overall architecture and key modules, and designed and implemented the price deviation monitoring logic.
Used Flink for real-time price validation, persisted live prices to Hive, and computed historical averages through scheduled jobs.
Triggered alerts and automatically filtered anomalous products during search, with a second blocking check at order placement.
Led backend work for a real estate platform spanning property data, new-home distribution, and financial reconciliation; the property data system handled about 100,000 QPS.
Designed and implemented order workflow and reporting modules, and participated in requirements analysis, prototype reviews, and UI reviews.
Drove GitFlow, DevOps, Scrum, and DDD practices; helped evolve a three-tier architecture toward DDD and extracted the housing module into an independent service.
Worked on the CRLIVE player, remote control, and distributed streaming services, developing and maintaining the player's long-running background service and streaming server components.
Contributed file management, upload/download, and voice-input features to CR Cloud Assistant; network communication used an internally designed and encapsulated protocol.
WPF, WinForm, FFmpeg, WebSocket, SRS
Guangdong Medical University
B.Eng. in Information Management and Information Systems · Sep 2013 — Jul 2017
Served as the primary programmer for the university proposal management system, implementing submission, review, analysis, progress tracking, and email notifications.
Earlier work also included a class-observation scheduling system, marking the transition from campus projects to full-time software engineering.
Writing & Open Source Technical Notes
I maintain long-running technical notes on software engineering, distributed systems, cloud-native infrastructure, incident analysis, and AI engineering. The blog and public repositories provide a broader record of my technical interests and growth beyond individual roles.