{"canonicalUrl":"https://humetric.net/p/profile-a20f09957c054e9090f3c28ff5d76fd4","profileVersion":"35c18dfb-fc67-4385-a5eb-54d11dcc188e","lastUpdated":"2026-10-03T16:05:15.790419+00:00","profile":{"display_name":"Wen-Chi (Vicky) Tsai","last_updated":"2026-10-03","slug":"profile-b31b57ffb49c4b97","headline":"AI Software Engineer","summary":"An AI software engineer focused on agent evaluation, model optimization, edge LLM compression, and multimodal machine learning. Experience includes building agent failure-analysis pipelines, MCP-based developer tools, quantized and distilled language models, geometric planning systems, and on-device computer vision applications.","years_experience":4,"skills":["A*","AI Agents","C++","Computer Vision","DFS","GitLab","GitLab Model Context Protocol","GPTQ","iOS","K-Nearest Neighbors","Knowledge Distillation","LLM Evaluation","LLM Fine-Tuning","LLM-Neo","LoRA","Machine Learning","MCP","Multimodal Machine Learning","Nearest-Neighbour Search","Open3D Point Cloud","Pattern Recognition","Python","PyTorch","Quantization","Rejection-Sampling Fine-Tuning","Sentence Transformers","Software Engineering","Swift","TensorFlow Lite","TypeScript","UAV Path Planning","VS Code","YOLOv8"],"experiences":[{"id":"1af[contact removed]a62-bd12-1d50e20896e9","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","title":"Research Assistant, Agent Harness and Model Co-Optimization","company":"Language Technologies Institute, CMU","start_date":"Sep. 2026","end_date":"Present","description":"Advised by Prof. Graham Neubig. Owned the data-selection track of a joint harness-and-model optimization project, turning failed agent trajectories into the training-data signal for rejection-sampling fine-tuning (RSFT). Built a failure-analysis pipeline over 720 agent trials of raw JSON telemetry, classifying each trial by exit state and exposing a 15.4% silent failure rate the reported mean absorbed, a 10-point AIME swing. Selected evaluation subsets with nearest-neighbour search over a 1.79M-row dataset, re-embedding all 17,398 DAPO-Math problems with sentence-transformers to rule out a text artifact, with 91 of 100 picks surviving.","position":0},{"id":"b93171ec-846b-4d97-bf6a-0d2bbd94126e","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","title":"Software Engineering Intern","company":"ASUS, Open Cloud Infrastructure Software R&D Team","start_date":"Jun. 2025","end_date":"Mar. 2026","description":"Automated issue triage with a GitLab Model Context Protocol (MCP) agent, pairing an LLM classification prompt with MCP tool calls to auto-classify, label, and route issues, then backfilling the manual backlog. Shipped four context providers into a VS Code AI coding assistant, adding commit-history, OS-metadata, current-file, and repository-map context selected from an evaluation of candidates. Cut a full regression pass from approximately 16 hours to approximately 1–2 hours with a TypeScript end-to-end (E2E) test suite covering approximately 95% of manual cases across Linux, Windows, and macOS.","position":1},{"id":"0b41ea2a-f2f8-4b59-ae15-543aa1f1a70a","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","title":"Research Assistant, Edge LLM Compression","company":"Wireless Ad-Hoc and Sensor Networks Laboratory, NCU","start_date":"Sep. 2025","end_date":"Feb. 2026","description":"Recovered 4-bit accuracy at 2-bit model size (IoU 0.9998, 1.18 GB) by fine-tuning a Llama-3.2 student with GPTQ quantization and LLM-Neo knowledge distillation through a LoRA adapter in PyTorch. Evaluated each configuration against 4-bit and 2-bit quantized baselines, tuning LoRA rank and KD loss weighting to isolate how much accuracy distillation recovered.","position":2},{"id":"108d5f98-acd2-4296-8e77-83f3c1f4c6b4","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","title":"Startup Engineer","company":"RoomMuse","start_date":"Jun. 2026","end_date":"Aug. 2026","description":"Caught unbuildable layouts at design time rather than on install day with a six-rule geometric conflict engine that blocks any unit sealing a socket, pipe, or doorway while passing intended connections, validated with 45 tests.","position":3}],"educations":[{"id":"ddfb8172-f[contact removed]bf9-b08bcd565a82","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","school":"Carnegie Mellon University (CMU), School of Computer Science, Language Technologies Institute","degree":"Master of Science","field":"Intelligent Information Systems","start_year":2026,"end_year":2027},{"id":"45826fa6-6726-4c25-89d0-82d426c8f4e3","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","school":"National Central University (NCU)","degree":"Bachelor of Science","field":"Computer Science and Information Engineering","start_year":2022,"end_year":2026}],"projects":[{"id":"361e[contact removed]f1b-9e2e-8df5df9cc010","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","name":"Museum Artifact Recognition App","description":"Shipped on-device model inference in an iOS app, running the team’s YOLOv8n and Pix2Pix reflection-removal models in Swift with TensorFlow Lite at 94.5% accuracy across 7 classes. Curated the 2,000-image training set and 699-image evaluation set spanning varied angles and lighting, and added reposition guidance so the app recovers from failed recognitions.","url":"https://github.com/vicky0619/museum-exhibition-classification","technologies":["iOS","Swift","TensorFlow Lite","YOLOv8n","Pix2Pix"],"position":0},{"id":"eb5186fc-8a9d-[contact removed]cc282017","person_id":"3684dab5-bda4-40fe-b70f-2134b64663fe","name":"Dual-Strategy UAV Path Planning for Building Structure Inspection","description":"Designed a dual-strategy planner trading layered MST-DFS for structured-space traversal against collision-aware A* replanning, cutting path-generation time over 5× and path length up to 14%. Built the 3D point-cloud pipeline in C++ and Python, slicing raw .pcd scans into layers and chaining per-layer KNN and MST into a DFS traversal, with normal-vector correction for full coverage.","url":"https://doi.org/10.1007/978-981-95-9843-4_21","technologies":["C++","Python","A*","KNN","MST","DFS","3D Point Clouds"],"position":1}]},"chunksVersion":"profile-chunks-v4","approvedDiscoveryQueries":[],"evidence":[{"id":"identity:0","type":"identity","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. 4 years experience","fieldKeys":["display_name","headline","years_experience"]},{"id":"summary:0","type":"summary","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Summary / 簡介: An AI software engineer focused on agent evaluation, model optimization, edge LLM compression, and multimodal machine learning. Experience includes building agent failure-analysis pipelines, MCP-based developer tools, quantized and distilled language models, geometric planning systems, and on-device computer vision applications.","fieldKeys":["summary"]},{"id":"skills:0","type":"skills","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Skills / 技能: A*, AI Agents, C++, Computer Vision, DFS, GitLab, GitLab Model Context Protocol, GPTQ, iOS, K-Nearest Neighbors, Knowledge Distillation, LLM Evaluation, LLM Fine-Tuning, LLM-Neo, LoRA, Machine Learning, MCP, Multimodal Machine Learning, Nearest-Neighbour Search, Open3D Point Cloud, Pattern Recognition, Python, PyTorch, Quantization, Rejection-Sampling Fine-Tuning, Sentence Transformers, Software Engineering, Swift, TensorFlow Lite, TypeScript, UAV Path Planning, VS Code, YOLOv8","fieldKeys":["skills"]},{"id":"discovery:0","type":"discovery","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Role / 職稱: AI Software Engineer. Experience / 年資: 4 years experience. Expertise / 專長: A*, AI Agents, C++, Computer Vision, DFS, GitLab, GitLab Model Context Protocol, GPTQ, iOS, K-Nearest Neighbors, Knowledge Distillation, LLM Evaluation","fieldKeys":["display_name","headline","years_experience","skills"]},{"id":"experience:0","type":"experience","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Experience / 經歷: Research Assistant, Agent Harness and Model Co-Optimization at Language Technologies Institute, CMU. Sep. 2026 to Present. Advised by Prof. Graham Neubig. Owned the data-selection track of a joint harness-and-model optimization project, turning failed agent trajectories into the training-data signal for rejection-sampling fine-tuning (RSFT). Built a failure-analysis pipeline over 720 agent trials of raw JSON telemetry, classifying each trial by exit state and exposing a 15.4% silent failure rate the reported mean absorbed, a 10-point AIME swing. Selected evaluation subsets with nearest-neighbour search over a 1.79M-row dataset, re-embedding all 17,398 DAPO-Math problems with sentence-transformers to rule out a text artifact, with 91 of 100 picks surviving.","fieldKeys":["experiences"]},{"id":"experience:1","type":"experience","ordinal":1,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Experience / 經歷: Software Engineering Intern at ASUS, Open Cloud Infrastructure Software R&D Team. Jun. 2025 to Mar. 2026. Automated issue triage with a GitLab Model Context Protocol (MCP) agent, pairing an LLM classification prompt with MCP tool calls to auto-classify, label, and route issues, then backfilling the manual backlog. Shipped four context providers into a VS Code AI coding assistant, adding commit-history, OS-metadata, current-file, and repository-map context selected from an evaluation of candidates. Cut a full regression pass from approximately 16 hours to approximately 1–2 hours with a TypeScript end-to-end (E2E) test suite covering approximately 95% of manual cases across Linux, Windows, and macOS.","fieldKeys":["experiences"]},{"id":"experience:2","type":"experience","ordinal":2,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Experience / 經歷: Research Assistant, Edge LLM Compression at Wireless Ad-Hoc and Sensor Networks Laboratory, NCU. Sep. 2025 to Feb. 2026. Recovered 4-bit accuracy at 2-bit model size (IoU 0.9998, 1.18 GB) by fine-tuning a Llama-3.2 student with GPTQ quantization and LLM-Neo knowledge distillation through a LoRA adapter in PyTorch. Evaluated each configuration against 4-bit and 2-bit quantized baselines, tuning LoRA rank and KD loss weighting to isolate how much accuracy distillation recovered.","fieldKeys":["experiences"]},{"id":"experience:3","type":"experience","ordinal":3,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Experience / 經歷: Startup Engineer at RoomMuse. Jun. 2026 to Aug. 2026. Caught unbuildable layouts at design time rather than on install day with a six-rule geometric conflict engine that blocks any unit sealing a socket, pipe, or doorway while passing intended connections, validated with 45 tests.","fieldKeys":["experiences"]},{"id":"education:0","type":"education","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Education / 學歷: Master of Science. Intelligent Information Systems. Carnegie Mellon University (CMU), School of Computer Science, Language Technologies Institute","fieldKeys":["educations"]},{"id":"education:1","type":"education","ordinal":1,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Education / 學歷: Bachelor of Science. Computer Science and Information Engineering. National Central University (NCU)","fieldKeys":["educations"]},{"id":"project:0","type":"project","ordinal":0,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Project / 專案: Museum Artifact Recognition App. Built with / 使用技術: iOS, Swift, TensorFlow Lite, YOLOv8n, Pix2Pix. Shipped on-device model inference in an iOS app, running the team’s YOLOv8n and Pix2Pix reflection-removal models in Swift with TensorFlow Lite at 94.5% accuracy across 7 classes. Curated the 2,000-image training set and 699-image evaluation set spanning varied angles and lighting, and added reposition guidance so the app recovers from failed recognitions.","fieldKeys":["projects"]},{"id":"project:1","type":"project","ordinal":1,"locale":"en","text":"Wen-Chi (Vicky) Tsai. AI Software Engineer. Project / 專案: Dual-Strategy UAV Path Planning for Building Structure Inspection. Built with / 使用技術: C++, Python, A*, KNN, MST, DFS, 3D Point Clouds. Designed a dual-strategy planner trading layered MST-DFS for structured-space traversal against collision-aware A* replanning, cutting path-generation time over 5× and path length up to 14%. Built the 3D point-cloud pipeline in C++ and Python, slicing raw .pcd scans into layers and chaining per-layer KNN and MST into a DFS traversal, with normal-vector correction for full coverage.","fieldKeys":["projects"]}]}