About Me

I am an AI systems architect and research-engineering leader.

I combine deep systems engineering with the ability to enter new domains quickly, identify the underlying abstractions, and turn ideas into working systems. Across my career, I have designed large-scale infrastructure, built AI products and models, developed novel research directions, and led the transition from early technical evidence to organizational adoption.

Changing Technical Direction Through Evidence

Much of my work has begun by questioning an assumption that constrained what an organization believed was possible. That pattern began in large-scale systems and later became central to my work in AI.

During Spanner’s early years, I designed comparative benchmarks against Bigtable and showed that, for representative read-heavy cloud workloads, stronger transactional guarantees did not necessarily require sacrificing performance. The work helped change discussion around when and how Spanner should be used.

At Google, I originated and developed DeepClue, a self-supervised representation-learning system that discovered semantic structure without predefined domain knowledge. I improved its performance by approximately 50× and adapted the same underlying representation across Google Search, Cloud, and Ads for integration-test selection, LLM context retrieval, and workflow optimization.

At Atlassian, I co-founded Rovo Dev CLI as a two-person project around the belief that the command line would become a critical interface for coding agents. When external open sourcing initially faced resistance, we built an internal open-source community instead. Within months, the product reached more than half of Atlassian engineers, achieved adoption comparable to Cursor internally, and helped establish the case for an external-facing product.

More recently, at Palona, I helped establish interaction tracking as a first-class perception target by building a lightweight model that raised accuracy from under 50% to over 80% while reducing per-frame latency from 15 seconds to 0.22 seconds.

These projects span distributed systems and AI, but they reflect the same working style: identify a consequential direction, challenge the assumption holding it back, build the evidence, and create the path from prototype to adopted capability.

Rethinking the Model–System Boundary

Many of the hardest AI engineering problems arise from how the model is separated from the system around it. Models are often treated as opaque endpoints, while reasoning, memory, verification, orchestration, and domain adaptation are built externally through prompts, generated tokens, and increasingly complex harnesses.

The same projects also reveal a second pattern: progress often depends on redesigning the boundary between the model and the system around it.

Rovo Dev CLI brought the agent into the engineer’s actual environment, where repositories, tools, tests, infrastructure, and organizational context were already present. DeepClue made a learned semantic representation reusable across distinct downstream systems. At Palona, Interaction Understanding aims to transform raw video into explicit representations of the relationships and state changes that operational software needs.

The data problem points in the same direction. Enterprises possess abundant proprietary code, documents, workflows, and video, but little labeled data capturing the concepts, interactions, and reasoning that matter. This makes concept representation learning a practical engineering requirement, not only a research preference.

My work on JEPA-style regularization explores how models can learn semantic structure at a coarser and more useful granularity than individual tokens. In Physical AI, Interaction Understanding asks the same question in the visual world: objects may be too fine and entire scenes too coarse, while interactions provide a more meaningful unit for prediction and action.

These efforts also motivate a broader view of reasoning. Rather than treating reasoning only as generated token sequences or a capability added during post-training, I am interested in how it can operate over learned concepts and internal state as an integral part of the model.

Toward Semantically Modular AI

Looking across these projects, an emerging theme is that AI models should become semantically modular gray-box components of software systems.

Future systems may need to connect software directly to learned concepts, internal representations, and model capabilities rather than interacting with models only through prompts and generated tokens. Such an architecture could make AI systems easier to integrate, adapt, verify, and debug. Above all, it could enable learning through reusable concepts—and that may be key to achieving human-level sample efficiency.