AI Systems Have Three Layers, Not Two
Writing
Production AI requires not only a model and a software system, but a reliable boundary that translates uncertain model outputs into usable software state.
Timeline
A chronological view of my work, research, products, and career milestones.
Built Palona's first interaction-tracking model
Builder
Built a lightweight model that established interaction tracking as a first-class perception target. Raised accuracy from under 50% with a flagship VLM to over 80%, while cutting per-frame latency from 15 seconds to 0.22 seconds.
Interaction Understanding: The Missing Layer for Physical AI
Writing
The physical world is not a collection of isolated objects. It is a stream of interactions. Interaction tracking is the right abstraction for Physical AI, yet it remains absent as a first-class perception target.
Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA
ICML 2026
Relaxes LLM-JEPA's two-view requirement through randomized fractional views and introduces a JEPA-style regularizer that improves signal-to-noise ratio and preserves diversity during LLM fine-tuning.
LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures
ICLR 2026
First successful integration of JEPA with LLMs, establishing JEPA as a practical regularizer for language models while eliminating the moving-average mechanism traditionally required to stabilize JEPA training.
Co-chaired AI/ML for Distributed Systems at ICDCS 2025
Service
Helped organize and shape the AI/ML for Distributed Systems track at ICDCS 2025.
Fine-tuned Atlassian’s first 70B code-review model
Builder
Led fine-tuning of Atlassian’s first 70B model for code review, using DeepSeek-generated review reasoning to improve review quality.
Adversarial Score identity Distillation: Rapidly Surpassing the Teacher in One Step
ICLR 2025
A one-step adversarial distillation method for diffusion models that improves both sample quality and distillation efficiency.
Co-founded Rovo Dev CLI
Builder
Originated the vision that the CLI would become the ultimate battleground for coding agents, despite the success of IDE-based workflows. Built a coding-agent CLI integrated with Atlassian’s Jira and Confluence ecosystem, bringing project context and developer workflows directly into the terminal.
Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation
ICML 2024
A data-free method that distills the generative capabilities of pretrained diffusion models into a single-step generator.
Developed DeepClue for self-supervised semantic discovery
Builder
Developed a self-supervised representation-learning system that discovered semantic structure without predefined domain knowledge. Improved system performance by approximately 50× and enabled applications across Google Search, Google Cloud, and Google Ads, including test selection, LLM context retrieval, and integration optimization.
Changed Google’s view of Spanner performance
Builder
Benchmarked Spanner against Bigtable and showed that, for representative read-heavy cloud workloads, Spanner could deliver lower latency while providing stronger transactional guarantees. The work helped shift Spanner’s perception from a specialized database accepted despite its performance tradeoffs, and also demonstrated approximately 3× higher simulated transaction throughput using three proposed capabilities.