Owned the evaluation and data-selection track of a research project that automatically improves LLM agents, building the Python tooling that tests each agent version and decides which problems it is measured on.
Found the project's real bottleneck was evaluation reliability: the benchmark disagreed with itself on 42% of problems under identical re-runs (Cohen's kappa 0.16), explaining three null experiments.
Ran the lab's self-improving agent loop end to end for the first time on the GPU cluster, fixing a breakage on main in a merged pull request so the system proposed its first new agent version from baseline failures.
Built a Python failure-analysis pipeline over 720 LLM agent runs, parsing raw JSON logs to expose a 15.4% silent failure rate the reported score had absorbed, a 10-point swing on the headline benchmark.
Kept evaluation running after external LLM APIs went down: self-hosted vLLM inference on parallel GPU servers via Slurm, setting concurrency from server queue metrics to cut projected runtime from 36 to 9 hours.
Built automated monitoring for long-running GPU jobs in Bash and Slurm that cancels a failing job within four minutes, after a silent grading misconfiguration scored 549 runs as zero with no error raised.
Cut a full regression pass from ~16 hours to ~1-2 hours with a TypeScript end-to-end (E2E) test suite covering ~95% of previously manual cases across Linux, Windows, and macOS.
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.
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.
Authored the assistant's Docusaurus documentation site, reducing hands-on onboarding support for adopting teams.
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 (45 tests).
Localized wall sockets from an iOS AR room scan built in Swift with RealityKit using a classical-CV pipeline (adaptive thresholding, connected components, NMS), ray-casting detections through the camera pose onto scanned wall planes.
Scaled teacher-student distillation to data-parallel training on 3 GPUs with DeepSpeed ZeRO-2, sharding optimizer state and gradients, since ZeRO-3 parameter sharding corrupts GPTQ-packed weights.
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.
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 in PyTorch.
Surveyed llama.cpp, vLLM, and SGLang for serving on a Jetson AGX Orin edge target, mapping latency versus throughput across PagedAttention, continuous batching, and prefix-cache reuse.
Authored the Vibe Coding column, translating AI coding workflows into hands-on tutorials that reached 50K+ readers and drove 42% platform content growth.
Designed and taught an AI-workflow course, plus the Her Path in Tech and Vibe Coding workshops, each reaching 200+ participants.
Frontend Engineer
Aug. 2024 - Jun. 2026
NCU College of Hakka Studies · Taoyuan, Taiwan
FrontendWeb PerformanceAccessibilitySecurity
Lifted mobile Core Web Vitals from 61% to 86% for 8,000+ monthly visitors, meeting WCAG 2.1 AA accessibility on a production site.
Remediated 17 security vulnerabilities across the codebase using automated review and scanning.
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.
Validated 100% collision-free operation in AirSim against a random-order baseline on real-world building models.
Museum Artifact Recognition App
May 2024 - Nov. 2024
SwiftTensorFlow LiteOn-Device InferenceiOSAWS EC2
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.
Integrated the iOS client with the team's AWS EC2 REST inference API, adding a cloud path for multi-user recognition.
Everything above is on one page, in ATS-friendly form, in my resume.