Caleb Kha-Uong

Dimare2026
Software Engineer Intern. Improved the shift scheduler for 3,300+ stores, including Starbucks Japan, cutting staffing gaps by 76%.
Spectre Intelligence (YC S26)2026
Member of Technical Staff. Built a forecaster of competing LLM agents that scores 0.153 Brier on ForecastBench.
Vision & Learning Lab2025
Researcher, with Jia Deng. Trained a Unitree G1 humanoid to walk to any goal and stop upright, and co-authored InFlux++.
Alexandria2025
Founding Engineer. Shipped search over a lab’s Slack, Drive, and papers that cites the source of every answer.
Intelligent Performance & Adaptation Lab2024
AI Researcher. Modeled false memories with semantic similarity, and studied why LLM agents fall into meltdown loops.

Bookworm
A lossless compressor for NASDAQ ITCH market data that replays the live order book as it codes, 2.33x smaller than zstd -19.
Unloop
Reproducing a rank-1 weight edit that stops Qwen3.5-0.8B thinking loops, tested against simpler interventions on held-out GSM8K and MATH.
World-Adapt
Six planning environments where the predictor stays wrong, and a notebook is the only memory between tasks. Paper.
FaceGS
Text-guided appearance editing for 3D Gaussian faces.

InFlux++
ECCV 2026. InFlux++: Real and Synthetic Data for Estimating Dynamic Camera Intrinsics
False Memories
CogSci 2026. Representational Similarity and Context Inference as a Shared Computational Account for False Memories in Humans

Previous Work