Robotics & Autonomous Systems
Intelligence
in the physical world
Nine years building AI software products has given me a systems perspective I believe is directly applicable to physical AI. This page is my public thinking about where I'm going next — and why.
Why Robotics
I'm transitioning intentionally into robotics and autonomous systems — not as a pivot away from AI product, but as its natural extension.
The product problems in physical AI are harder, the stakes are higher, and the impact is more tangible. Every principle I've developed in search, ranking, and LLM products — evaluation rigor, feedback loop design, uncertainty handling, human-in-the-loop systems — applies directly to autonomous systems product management.
I'm studying, building perspective, and actively seeking roles and collaborations at the frontier of physical AI product development.
Areas of Focus
Four domains I'm studying
Autonomous Driving
AD systems are the most complex product problem of our generation: real-time decision-making under uncertainty, multi-modal sensor fusion, and a safety envelope that tolerates zero ambiguity. The product discipline required to ship safe AV systems will define the next era of product management.
- Sensor fusion architecture and perception stack design
- Behavioral planning under uncertainty and edge cases
- Safety-critical evaluation frameworks
- Operational domain expansion strategy
Robotics Product Management
Physical AI product management combines the rigor of embedded systems with the user-centricity of consumer software. The unique challenge: your product breaks in the real world, in real time, with real consequences — and you can't roll back a hardware deployment.
- Hardware-software co-design for productization
- Teleoperation to autonomy transition strategy
- Robot learning from human demonstration
- Fleet management and operational scaling
AI Agents in Physical Space
The same agent design principles from software — goal decomposition, tool use, state tracking, recovery from failure — apply in robotics, but with latency constraints and physical consequences. This is where my LLM product background connects directly to physical AI.
- Task planning and execution in unstructured environments
- Failure recovery and graceful degradation design
- Human-robot interaction and trust calibration
- Sim-to-real transfer and evaluation methodology
Embodied Intelligence
Embodied AI represents a fundamental shift: intelligence that exists in and acts on the physical world. Understanding how context, affordances, and physical constraints shape intelligent behavior is the defining product challenge of physical AI development.
- Foundation models for robotic manipulation
- Affordance understanding and scene comprehension
- Dexterous manipulation and fine motor control
- Long-horizon task planning with world models
Transferable Expertise
From software AI to physical AI
Evaluation frameworks for LLMs
Evaluation frameworks for robotic behaviors
Human-in-the-loop LLM systems
Human-in-the-loop robot teleoperation
A/B experimentation at scale
Sim-to-real transfer validation
Feedback loop design for ranking
Reward signal design for RL
Cold start problems in search
Limited data problems in robot learning
Agent orchestration patterns
Task planning for autonomous systems
Active Reading
Scaling Laws for Neural Language Models
Kaplan et al.
RT-2: Vision-Language-Action Models
Google DeepMind
Do As I Can, Not As I Say
Google Research
Open X-Embodiment
Google DeepMind
Physical Intelligence — Pi0
Physical Intelligence
Foundation Models for Decision Making
Yang et al.
Building something at the frontier of physical AI?
Let's talk