Backend / Systems / Applied AI Engineer
Shufeng Chen
M.S. Electrical Engineering at Columbia University. My work spans cloud-native services, distributed systems, and AI product engineering — with an emphasis on reliability, observability, and clean API design.
Shufeng (Alex) Chen
Columbia University · M.S. EE
What I Build
Focus Areas
Cloud Backend
Microservices, REST APIs, caching, data pipelines. Building scalable cloud-native services on Kubernetes with emphasis on reliability and observability.
Applied AI
LLM applications, retrieval-augmented generation, ML experimentation and evaluation. Turning research into production-grade AI products.
Systems
Performance debugging, reliability engineering, and dev tooling. Distributed consensus, concurrency control, and low-level optimization.
Career Journey
Experience
Software Engineer Intern
Tencent · Cloud Computing Group
- Built backend microservices for distributed inference pipelines on Kubernetes-managed HPC clusters; automated container lifecycle & remote job operations for smoother orchestration.
- Developed internal cluster management and scheduling APIs (Spring Boot/Security, Redis/MySQL) and integrated cloud metadata synchronization for lifecycle tracking.
- Improved ingestion/telemetry services by shipping a CLI-based pipeline adopted by 20+ internal apps; enabled high-throughput streaming from 4,300+ edge nodes.
- Deployed pre-embedding pipelines across Tencent Cloud and AWS (EKS/EC2), improving inference latency by 1.3× and reducing CPU load by 2.2× via caching and preprocessing.
Full Stack Software Engineer Intern
Tree-Graph Research Institute
- Built a cross-platform crypto analytics app (React Native) with GPT-based insights; summarized 5,000+ news sources using LDA topic modeling.
- Designed an automated ETL pipeline (Scrapy, AWS Lambda, DynamoDB, EventBridge) and indexed real-time data to OpenSearch; improved search API performance by 33%.
- Shipped high-performance secrets-detection microservices (Go/Python/Rust) over gRPC + Kubernetes; reduced P95 latency from 170ms to 40ms and supported 10K+ QPS using Bloom-filter warm cache.
- Implemented serverless remediation workflows (Step Functions + Lambda) and production observability (Prometheus/Grafana, structured logs).
Full Stack Software Engineer Intern
VisionX LLC
- Developed an ONNX-optimized, real-time environmental monitoring platform with depth-enabled cameras and on-device inference for fire/smoke and pedestrian detection.
- Led delivery of a PDF Q&A system on AWS (SageMaker/App Runner/S3) with LangChain + React + Express; improved responsiveness via vector caching and prompt tuning.
- Designed an MLOps rollout pipeline (offline → shadow → live) with Jenkins and Grafana for drift/latency monitoring in field deployments.
QuantHarbor
End-to-end AI financial research platform transforming unstructured market documents into citation-grounded insights.
- Multi-Agent Architecture: Orchestrated specialized agents (Data Collector, Analyzer, Report Generator, Deep Search) collaborating in a shared variable space.
- RAG Pipeline: Built a document-to-insight pipeline integrating PDF ingestion, vector indexing, and citation tracing for verifiable insights.
- VLM Feedback Loops: Implemented built-in vision agents that automatically correct chart issues during publication-grade report generation.
- Interactive UI: Delivered a full-stack dashboard for system configuration, execution monitoring via WebSockets, and checkpoint/resume management.
Distributed Key-Value Store
Horizontally scalable, fault-tolerant key-value storage system featuring strong consistency and distributed transactions.
- Raft Consensus & Sharding: Implemented Multi-Raft for log replication, dynamic region splitting, and leader transfer without downtime.
- Global Scheduler: Developed a heartbeat-driven scheduler to monitor cluster metadata and auto-rebalance replicas across nodes.
- Percolator 2PC: Built distributed transaction support with Snapshot Isolation handling prewrites, commits, and rollbacks.
- Concurrency Control: Managed concurrent operations via MVCC and per-key latching on an embedded LSM-tree engine.
Patton Food
Full-stack, cloud-deployed online food ordering web application with clean API design and a modern UI.
- User registration and login with session-based authentication using Spring Security.
- Password encryption with BCrypt and secure session management.
- Restaurant/menu browsing with cart and checkout flow.
- Cloud deployment on AWS with production-ready configuration.
PDF-AI: Conversational Q&A System
AI assistant that lets users upload PDFs and ask natural language questions about their content.
- PDF upload & processing pipeline with chunking and vector indexing.
- Retrieval + LLM answering with conversation context and source citations.
- Responsive chat UI with streaming responses.
- Latency optimizations via caching and request shaping.
L2 Game Engine
Cross-platform 2D game engine in C++ with Lua scripting — designed to keep the runtime fast, modular, and easy to extend.
- C++ core runtime: real-time game loop, scene/actor system, and engine APIs optimized for iteration speed.
- Lua scripting + language hosting: externalized gameplay logic with script-facing APIs for actor creation and behaviors.
- Box2D physics: collision handling, rigid body dynamics, friction, and gravity for realistic 2D gameplay.
- Shipped a complete game on the engine: Havoc: 300-Seconds Escape.
LinguAR: AR Language Learning
iOS AR app prototype combining real-scene recognition, translation, and multimodal tutoring for daily language practice.
- Built the end-to-end AR learning loop: object recognition (YOLOX) → translation overlay (Google Cloud Translate) → pronunciation output (AWS Polly).
- Led AR shooting mini-game feature implementation: fixed screen flicker, added SFX/VFX + scoreboard.
- Integrated voice-first AI tutoring (Whisper STT + GPT-4o + TTS) for conversational practice with multimodal I/O.
- Delivered a usability-tested team prototype with story maps and iterative refinements.
Interactive Work
Game & XR Projects
Unreal Engine · Blueprints
VR walkthrough simulator focused on interaction and environment fidelity, built for real-time performance.
- Interactive VR navigation with user-friendly scene flow
- Blueprint-driven interaction logic for rapid iteration
- Performance-oriented real-time rendering considerations
Unity · C# · Mapbox
AR experience inspired by location-based gameplay, combining map navigation with AR interactions.
- Map-based discovery using Mapbox integration
- AR interactions and UI flows for collection/progression
- Mobile-first performance and interaction design
Unity · XR/AR · Meta Quest
AR prototype for previewing furniture in space before purchase, emphasizing scale, placement, and usability.
- Virtual placement workflow for in-room visualization
- Interaction and UI patterns optimized for XR devices
- End-to-end prototype delivery with structured iteration
Unity · C#
Unity remaster-style gameplay prototype focusing on core action-adventure mechanics and moment-to-moment feel.
- Implemented key gameplay systems (combat/items/progression)
- Enemy behavior and encounter tuning for gameplay clarity
- Polished controls and UI feedback loops
Unity · C#
Gameplay prototype built around echolocation-style exploration, turning sound-driven feedback into navigation.
- Core mechanic: reveal environment state through echolocation cues
- Designed exploration loop to support discovery and tension
- Implemented interaction and pacing to keep sessions engaging
Unity · C#
Puzzle game centered on portal mechanics, emphasizing level logic, item interactions, and UI clarity.
- Portal-based puzzle design with staged difficulty progression
- Inventory UI and interaction system for puzzle tooling
- Shipped playable build with multiple levels and clear objectives
Research
Publications
Breast Cancer Detection Model Training Strategy Based on Continual Learning
CAIBDA 2022 · Conference Paper
Shufeng Chen, Fenghui Tang.
IEEE XploreGlobal-local Dynamic Adversarial Learning for Cross-domain Sentiment Analysis
Mathematics 11(14):3130 · 2023
Juntao Lyu, Zheyuan Zhang, Shufeng Chen, Xiying Fan*.
MDPI