{"canonicalUrl":"https://humetric.net/p/profile-73c4368670b54e6e916afe0f284da625","profileVersion":"2bd98ede-64b0-4c8d-aa56-7e2a7b5a3e1a","lastUpdated":"2026-10-19T14:45:43.774197+00:00","profile":{"display_name":"Winona Chu","last_updated":"2026-10-19","slug":"profile-1b5ea067e74942c7","headline":"Software Engineer","summary":"A full stack software engineer and machine learning researcher with experience building distributed systems, AI-powered applications, computer vision pipelines, and cloud platforms. Their work spans AWS, backend and frontend development, anomaly detection, multimodal models, mixed reality, and fault-tolerant systems.","years_experience":1,"skills":["Amazon RDS","Amazon S3","AWS","AWS Bedrock","AWS CDK","AWS Lambda","AWS Step Functions","Bash","C/C++","C#","Claude Code","Codex","Docker","FastAPI","Git","GitHub Actions","Hugging Face","Java","JavaScript","JavaScript/TypeScript","MCP","NumPy","OpenAI API","Pandas","PostgreSQL","PostHog","Python","PyTorch","React","Scikit-learn","SLURM","SQL","TensorFlow","Tool Calling","TypeScript"],"experiences":[{"id":"ba25a522-efd5-4894-a3e7-822ad65272a3","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","title":"Software Engineer","company":"KhoriumAI","start_date":"2026/06/13","end_date":"present","description":"● Reduced billed LLM context by 63% (4,022 → 1,497 tokens/request) by moving an 85-tool MCP catalog behind on-demand ToolSearch and prompt caching, improving agent efficiency across prod application on AWS Bedrock.\n● Prevented 7–10 minutes of wasted meshing on invalid manufacturing CAD by building an asynchronous\npre-mesh geometry validation workflow with bounded execution and automatic fallback.\n● Reduced duplicate container builds by 67% by replacing three environment-specific builds with one immutable build per commit, promoting identical artifacts through development, preproduction, and production.\n● Eliminated inconsistent engineering calculations across the UI, AI agent, and generated reports by consolidating duplicated logic into a single backend assembly graph for CFD/CHT/FEA, with closure-gap detection, symbolic solving, and validation of agent-composed equations.","position":0},{"id":"667a1184-3d[contact removed]ba-60e1a146cbf9","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","title":"Full Stack Software Engineer","company":"astrolabe analytics (UW Capstone)","start_date":"2025/01/01","end_date":"2025/06/12","description":"● Enabled multi-tenant ingestion and querying across 6 battery-data formats by building a React/FastAPI platform with 15+ REST endpoints and normalized PostgreSQL schemas backed by AWS RDS and S3.\n● Raised schema column matching accuracy to 90% by building a 3-layer standardization pipeline using\nFuzzyWuzzy, OpenAI API, and ISO 80000 unit conversion.","position":1},{"id":"029ab262-9c90-49dc-b3ec-f039f63f6afd","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","title":"Software Development Engineer Intern","company":"Innodisk Corporation","start_date":"2022/07/01","end_date":"2022/08/31","description":"● Expanded training data 5.6× (1.3k → 7.6k images) for a real-time screw-welding anomaly detection pipeline using geometric and noise-based augmentation.\n● Achieved 100+ FPS at 8W with 90% mAP@0.45 on a Xilinx KV260 by fine-tuning YOLOv4-tiny in Keras and deploying an int8-quantized model through Vitis-AI (.h5 → q.pb → .xmodel).","position":2}],"educations":[{"id":"7e5f7d14-6f8e-4116-81be-707173afe62d","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","school":"University of Washington","degree":"Master's","field":"Electrical and Computer Engineering","start_year":2024,"end_year":2026},{"id":"32d3636b-271c-4d15-ba4e-09c6e4cc8641","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","school":"National Chengchi University","degree":"Bachelor's","field":"Management of Information Systems","start_year":2020,"end_year":2024}],"projects":[{"id":"67e38238-e562-4f05-a446-e181a26afa28","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","name":"Machine Learning Research Assistant at Information Processing Lab (UW)","description":"Achieved 95.62% Pixel AUC on zero-shot remote-sensing anomaly detection by co-authoring a CVPR 2026 Workshop project using CLIP + DINOv3 dual-vision fusion. Scaled semiconductor-defect VQA training to a 172K+ image dataset by building distributed SLURM pipelines and fine-tuning Qwen3.5-VL with GRPO and custom reward functions to reduce hallucinations. Cut manual annotation effort by 50% and generated 50K+ labeled images by building an end-to-end GroundingDINO + SAM2 auto-annotation pipeline.","url":null,"technologies":["CLIP","DINOv3","SLURM","Qwen3.5-VL","GRPO","GroundingDINO","SAM2"],"position":0},{"id":"7fb28f0a-bf3e-4de6-91bb-ba630a933967","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","name":"Sharded Key-Value Distributed System","description":"Enabled linearizable multi-key transactions and dynamic re-sharding across server groups by building a fault-tolerant sharded key-value store in Java using Two-Phase Commit (2PC). Maintained replica consistency under server crashes and lossy networks by implementing Multi-Instance Paxos and deterministic message handlers.","url":null,"technologies":["Java","Two-Phase Commit","Multi-Instance Paxos"],"position":1},{"id":"60bd3e8c-46ad-4d18-8f5f-b0736544cdf3","person_id":"ae8582cc-9926-4ecb-9666-681ee054fbd3","name":"HoloLights: Color Vision Deficiency MR App","description":"Won 1st Place in the 2023 National Universities Innovation Competition by building a Unity/C# mixed-reality accessibility app for HoloLens 2 with real-time color transformation. Reduced edge-to-server synchronization latency by 100ms using sockets and C# coroutines while integrating TensorFlow object detection with 94% mAP.","url":"https://winona1111.github.io/UWCapstone-project/","technologies":["Unity","C#","HoloLens 2","TensorFlow","Sockets"],"position":2}]},"chunksVersion":"profile-chunks-v4","approvedDiscoveryQueries":[],"evidence":[{"id":"identity:0","type":"identity","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. 1 years experience","fieldKeys":["display_name","headline","years_experience"]},{"id":"summary:0","type":"summary","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. Summary / 簡介: A full stack software engineer and machine learning researcher with experience building distributed systems, AI-powered applications, computer vision pipelines, and cloud platforms. Their work spans AWS, backend and frontend development, anomaly detection, multimodal models, mixed reality, and fault-tolerant systems.","fieldKeys":["summary"]},{"id":"skills:0","type":"skills","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. Skills / 技能: Amazon RDS, Amazon S3, AWS, AWS Bedrock, AWS CDK, AWS Lambda, AWS Step Functions, Bash, C/C++, C#, Claude Code, Codex, Docker, FastAPI, Git, GitHub Actions, Hugging Face, Java, JavaScript, JavaScript/TypeScript, MCP, NumPy, OpenAI API, Pandas, PostgreSQL, PostHog, Python, PyTorch, React, Scikit-learn, SLURM, SQL, TensorFlow, Tool Calling, TypeScript","fieldKeys":["skills"]},{"id":"discovery:0","type":"discovery","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. Role / 職稱: Software Engineer. Experience / 年資: 1 years experience. Expertise / 專長: Amazon RDS, Amazon S3, AWS, AWS Bedrock, AWS CDK, AWS Lambda, AWS Step Functions, Bash, C/C++, C#, Claude Code, Codex","fieldKeys":["display_name","headline","years_experience","skills"]},{"id":"experience:0","type":"experience","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. Experience / 經歷: Software Engineer at KhoriumAI. 2026/06/13 to present. ● Reduced billed LLM context by 63% (4,022 → 1,497 tokens/request) by moving an 85-tool MCP catalog behind on-demand ToolSearch and prompt caching, improving agent efficiency across prod application on AWS Bedrock.\n● Prevented 7–10 minutes of wasted meshing on invalid manufacturing CAD by building an asynchronous\npre-mesh geometry validation workflow with bounded execution and automatic fallback.\n● Reduced duplicate container builds by 67% by replacing three environment-specific builds with one immutable build per commit, promoting identical artifacts through development, preproduction, and production.\n● Eliminated inconsistent engineering calculations across the UI, AI agent, and generated reports by consolidating duplicated logic into a single backend assembly graph for CFD/CHT/FEA, with closure-gap detection, symbolic solving, and validation of agent-composed equations.","fieldKeys":["experiences"]},{"id":"experience:1","type":"experience","ordinal":1,"locale":"en","text":"Winona Chu. Software Engineer. Experience / 經歷: Full Stack Software Engineer at astrolabe analytics (UW Capstone). 2025/01/01 to 2025/06/12. ● Enabled multi-tenant ingestion and querying across 6 battery-data formats by building a React/FastAPI platform with 15+ REST endpoints and normalized PostgreSQL schemas backed by AWS RDS and S3.\n● Raised schema column matching accuracy to 90% by building a 3-layer standardization pipeline using\nFuzzyWuzzy, OpenAI API, and ISO 80000 unit conversion.","fieldKeys":["experiences"]},{"id":"experience:2","type":"experience","ordinal":2,"locale":"en","text":"Winona Chu. Software Engineer. Experience / 經歷: Software Development Engineer Intern at Innodisk Corporation. 2022/07/01 to 2022/08/31. ● Expanded training data 5.6× (1.3k → 7.6k images) for a real-time screw-welding anomaly detection pipeline using geometric and noise-based augmentation.\n● Achieved 100+ FPS at 8W with 90% mAP@0.45 on a Xilinx KV260 by fine-tuning YOLOv4-tiny in Keras and deploying an int8-quantized model through Vitis-AI (.h5 → q.pb → .xmodel).","fieldKeys":["experiences"]},{"id":"education:0","type":"education","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. Education / 學歷: Master's. Electrical and Computer Engineering. University of Washington","fieldKeys":["educations"]},{"id":"education:1","type":"education","ordinal":1,"locale":"en","text":"Winona Chu. Software Engineer. Education / 學歷: Bachelor's. Management of Information Systems. National Chengchi University","fieldKeys":["educations"]},{"id":"project:0","type":"project","ordinal":0,"locale":"en","text":"Winona Chu. Software Engineer. Project / 專案: Machine Learning Research Assistant at Information Processing Lab (UW). Built with / 使用技術: CLIP, DINOv3, SLURM, Qwen3.5-VL, GRPO, GroundingDINO, SAM2. Achieved 95.62% Pixel AUC on zero-shot remote-sensing anomaly detection by co-authoring a CVPR 2026 Workshop project using CLIP + DINOv3 dual-vision fusion. Scaled semiconductor-defect VQA training to a 172K+ image dataset by building distributed SLURM pipelines and fine-tuning Qwen3.5-VL with GRPO and custom reward functions to reduce hallucinations. Cut manual annotation effort by 50% and generated 50K+ labeled images by building an end-to-end GroundingDINO + SAM2 auto-annotation pipeline.","fieldKeys":["projects"]},{"id":"project:1","type":"project","ordinal":1,"locale":"en","text":"Winona Chu. Software Engineer. Project / 專案: Sharded Key-Value Distributed System. Built with / 使用技術: Java, Two-Phase Commit, Multi-Instance Paxos. Enabled linearizable multi-key transactions and dynamic re-sharding across server groups by building a fault-tolerant sharded key-value store in Java using Two-Phase Commit (2PC). Maintained replica consistency under server crashes and lossy networks by implementing Multi-Instance Paxos and deterministic message handlers.","fieldKeys":["projects"]},{"id":"project:2","type":"project","ordinal":2,"locale":"en","text":"Winona Chu. Software Engineer. Project / 專案: HoloLights: Color Vision Deficiency MR App. Built with / 使用技術: Unity, C#, HoloLens 2, TensorFlow, Sockets. Won 1st Place in the 2023 National Universities Innovation Competition by building a Unity/C# mixed-reality accessibility app for HoloLens 2 with real-time color transformation. Reduced edge-to-server synchronization latency by 100ms using sockets and C# coroutines while integrating TensorFlow object detection with 94% mAP.","fieldKeys":["projects"]}]}