Homer Quan

I live in New York City with my wife and our two children. They are the fixed point in a life that has crossed countries, companies, ideas, and more than a few uncertain experiments. Work matters to me, but it is not the center of the story. Home is.

That perspective shapes how I approach technology. The objective is not intelligence in isolation, but systems that strengthen human agency and produce meaningful outcomes in the environments where people actually work.

My work is driven by a simple belief: AI should help people think more clearly, learn faster, coordinate better, and act with confidence in complex environments. The chapters that follow trace how that belief has taken form in knowledge systems, products, teams, and machines.

A recurring theme in my work is the representation of knowledge: how entities, events, rules, evidence, and human intent can be organized into a model of a domain. At Yahoo, this took the form of graph entity search, large-scale graph data pipelines, and language-driven content analysis. The same foundations support applications such as intelligent analysis, content integrity, and fake-news detection, where isolated signals matter less than their provenance and relationships.

Knowledge graphs and ontologies provide explicit structure; neural models contribute perception and pattern recognition; symbolic rules add constraints and explanations. I am particularly interested in this neurosymbolic combination because it supports reasoning that is both capable and auditable.

At Wise Systems, that structure became a real-time decision engine for routing vehicles through live traffic, telemetry, predictions, and operational constraints. At Reflen, it represented expert behavior and feedback. The pattern is similar to operational decision platforms such as Palantir AIP: connect domain ontology, live data, models, policy, and human approval so decisions can be understood and acted upon.

Much of my work begins at zero: an ambiguous problem, a technical possibility, and no established path between them. I translate that possibility into an architecture, build the highest-risk components, connect them into an application, and establish the operational foundations required for production.

The progression moves from concept and prototype to product and platform. It may involve an agent workflow, a graph service, a voice pipeline, a simulator bridge, or an end-to-end application. The specific component matters, but so do the interfaces between components—the places where data, models, users, and operational constraints meet.

Building also means creating the culture and team around the system. Clear technical principles, shared ownership, direct feedback, and disciplined experimentation allow a small team to move quickly without sacrificing reliability. My role is hands-on: clarify the problem, make the first version real, and help the team develop the judgment to carry it forward.

Pajama Cats began close to home: with children, books, and the possibility that a story might listen as well as speak. We built voice, language-model, content, and mobile components into an interactive application while treating age-appropriate boundaries, privacy, validation, refusals, and parent controls as product requirements.

Later, the setting changed to enterprise compliance. Policies, structured records, unstructured documents, communication history, retrieval, and model reasoning had to converge in a workflow that preserved reviewer authority. Across both contexts, the principle was consistent: AI may propose, summarize, and reason, but consequential decisions need traceability, explicit constraints, and a clear path for human intervention.

At SP8CEAI, I led the architecture of human-robot workflows for space and industrial operations. The physical AI stack connected language-model agents, procedure understanding, symbolic checks, ROS execution, Omniverse simulation, AR/VR review, and operator approval into an auditable path from plan to action.

On-edge AI introduces a different set of engineering constraints: inference latency, intermittent connectivity, device coordination, privacy, resilience, limited compute, and safe failure modes. It requires more than deploying a model to hardware. The surrounding infrastructure must manage tasks, state, model routing, checkpoints, observability, and human control across local devices, robots, simulators, and private compute.

MirrorNeuron explores the software side of that problem. Its local-first runtime treats retries, approvals, sleep and resume, and portable execution as core primitives. Together, these efforts point toward AI systems that can operate close to the work while remaining inspectable, recoverable, and governed by the people responsible for the outcome.