I-Ju Lin

I'm a software engineer at Klaviyo (🇺🇸), on the AI & Analytics Advanced Reporting team, where I work on Snowflake analytics pipelines and Kubernetes-containerized Spark jobs orchestrated through Airflow on AWS.

Before Klaviyo I spent two years at Cognex shipping machine vision products from zero to one — Yocto embedded Linux, gRPC microservices, and high-performance C++ inter-process communication. Along the way I was the founding AI engineer at KardoAI (NVIDIA Inception Program), building a FinTech LLM agent with tool-calling and Mastercard API integrations, and interned at Dell on bare-metal telemetry for telecommunications. I started my career at 104 Corporation in Taiwan (🇹🇼), building a TypeScript SDK for OAuth and OpenID Connect.

I hold an M.Sc. in Computer Science from Boston University, where I served as International Student Relations Chair of the Graduate Student Organization, and an M.Sc. & B.S. in Computer Science from National Tsing Hua University. My master's thesis was on self-supervised robotic control using reinforcement learning, and I published human-computer interaction research at CHI. These days I'm most excited about continual learning on small, on-device language models.

news

May 2026Started Online-SDFT: on-device continual learning from hindsight — a small language model rereads its own decisions once outcomes arrive and distills that hindsight into a LoRA adapter, running on Android. Prototyped with Zhang-Wei Hong and Kai-Chi Huang at the Sundai Hackathon.

Mar 2026Joined Klaviyo's AI & Analytics Advanced Reporting team in Boston.

Jun 2024Joined Cognex Corporation as a Software Engineer working on machine vision and embedded systems.

Jan 2024Graduated from Boston University with an M.Sc. in Computer Science.

Apr 2020Our paper on smart-home routine assistants was presented at CHI 2020.

publications

Exploring the Design Space of User-System Communication for Smart-home Routine Assistants

Yi-Shyuan Chiang, Ruei-Che Chang, Yi-Lin Chuang, Shih-Ya Chou, Hao-Ping Lee, I-Ju Lin, Jian-Hua Jiang Chen, and Yung-Ju Chang

CHI Conference on Human Factors in Computing Systems (CHI '20), Honolulu, HI, USA, 2020

A user-enactment study of how AI-enabled smart-home agents that automate household routines should communicate with users — how people interpret a system's confidence in its automated acts, how they assess, diagnose, and improve them, and design recommendations for the user-system communication design space of smart-home routine assistants.

DOI PDF

projects

SLM-Online-SDFT — continual on-device fine-tuning for small language models: on-policy self-distillation with no labels and no external teacher, distilling hindsight into a lightweight LoRA adapter on a LiquidAI LFM2.5-230M model. Includes a notification-routing benchmark, a Colab notebook, and an Android demo.

AI-based Prompt Optimizer — integrated Group Relative Policy Optimization (GRPO) and GEPA (Genetic-Pareto) into a ReAct agent framework with KL coefficients and PPO clipping for reward modeling; trained a Qwen3-8B inference server with a retrieval node using Fully Sharded Data Parallel, on vLLM, VeRL, Ray, and Slurm.

Boston Marathon Race Simulator — interactive race simulation with weather-driven pacing, gear recommendations, and a rule-based virtual coach.

Extension for Web Development — published an npm package for BrowserStack version compatibility and extended TypeScript support in JavaScript libraries: CORS, cookies stored by Redis, and web sessions.