Skip to main content

In Keun Kim

About

I am an M.S. student in Computer Science at Columbia University, advised by Professor Baishakhi Ray.

I am currently interested in trustworthy AI for software engineering, particularly cybersecurity software agents and methods for validating AI-generated software changes.

Before Columbia, I spent more than six years as a software engineer, primarily at Toss, a financial super app used by more than half of South Korea’s population.

I built production systems for real-time fraud detection, personalized advertising recommendation and inference, and unified multi-model LLM serving, among others.

Outside research, I enjoy
scotch whisky, specialty coffee, and cats.

Research

  1. [1] Elaine Ang, Sam Weldon, In Keun Kim, Kevin Durand, Kostis Kaffes, and Eugene Wu, “BranchBench: Aligning Database Branching with Agentic Demands,” arXiv preprint, 2026. [arXiv]

Professional Experience

  • Personalized Advertisement Inference Platform

    Built and operated a production deep-learning inference path that served personalized ads at over 6,000 requests per second with p95 latency under 70 milliseconds, including traffic splitting for online experiments and hourly model rollout without downtime.

    Led an architecture redesign that cut more than 1,500 lines of duplicated serving logic, reduced pod memory by 25 percent, and shortened new API delivery from about a day to under three hours—making it practical to ship and validate model and ranking changes on a tight loop.

  • Real-Time Fraud Detection System

    Owned real-time fraud scoring and transaction blocking that protected users under strict latency budgets (p95 ≈ 200 ms, p99.9 ≈ 600 ms) while scoring against more than 1 TB of live features—turning model decisions into reliable, low-latency controls in a high-stakes financial environment.

  • LLM Serving Platform

    Built a unified platform for serving in-house and hosted language models with shared request interfaces, retrieval over internal knowledge, and batching—so product teams could evaluate and ship model-backed features without each owning a separate serving stack.