TS
Portrait of Teymur Sadikhov
Fig. 01 San Jose, California depth

Physical AI · Agentic Systems · AI Infrastructure

Dr. Teymur
Sadikhov

I build AI that runs in production, in companies and in machines.

Twenty years of AI, starting with a paper on reinforcement learning for multi-agent systems. I build agentic systems that run real enterprise workflows, the AI infrastructure that trains and serves them, and the physical AI that puts models onto robots and vehicles.

Read the trajectory Work with me
Now
VP, AI & Physical AI
In AI and robotics
20 years
Doctorate
AI & Robotics

01 · Origins

It started with multi-agent RL

My first paper was on reinforcement learning for multi-agent systems. That was nearly twenty years ago, and it is still more or less what I work on.

Since then I have put AI onto most classes of machine that move: autonomous helicopters and drones, unmanned ground vehicles on and off road at NASA JPL, self-driving cars at Mercedes-Benz, Cyngn and XPeng, warehouse robots at Volax, and lately robotic arms and humanoids running vision-language-action models.

The formal training is aerospace. I did my doctorate at Georgia Tech and a master's at UC San Diego, both on AI and robotics, and I still use that side of it when a client wants autonomy on aircraft or drones rather than on the ground. Before either of those I studied aeronautical engineering at Istanbul Technical University and graduated as the top student of the university.

The multi-agent framing never went away. Twenty years ago it meant deriving coordination protocols by hand for a handful of vehicles. Now it means foundation models and fleets of software agents, and the open questions are the same ones: who decides what, and what happens when one of them is wrong.

02 · Work

Four things people call me about

These were separate jobs for most of my career. In the last three years they stopped being separate, and I usually end up working on some combination of all four at once.

  1. i

    AI infrastructure

    AI Factories: the GPU cluster, the orchestration layer, and the platform teams actually build on. I put together Deloitte's turnkey version, running multi-node training on H100, H200 and Blackwell systems with NeMo and NIM under Slurm and Kubernetes. Clients see roughly 50% lower total cost of ownership. A good share of it runs in sovereign environments where the data cannot leave the country, which constrains the architecture in ways that are easy to underestimate early on.

  2. ii

    Enterprise AI agents

    Agents that take on repeatable operational work rather than answering questions in a chat window. In practice they sit against ERP, CRM and support systems and run governed workflows across operations, finance, logistics and customer care, with audit trails and a draft-and-approve step so a person signs off and access stays scoped by role. Underneath sits a context layer over the company's own records that answers in plain language and cites its sources.

  3. iii

    Physical AI

    Vision-language-action models on real hardware, mostly robotic arms and humanoids. I have run pick-and-place and dexterous manipulation programs with Nvidia Isaac GR00T and Cosmos, trained in Isaac Sim and deployed to Jetson Thor for onboard inference. Cross-environment generalization is what I test first. Most demos are tuned to one room, and they come apart as soon as the lighting or the fixture changes.

  4. iv

    Robotics and autonomous vehicles

    Twenty years of autonomy across air and ground: helicopters and drones, unmanned vehicles on and off road, passenger cars, and warehouse fleets. I have led the AI and autonomy stack at three vehicle companies, and I now advise on end-to-end sensor-to-action policies, pretraining infrastructure and simulation strategy.

03 · Thesis

Where the difficulty actually sits

Almost every team I meet thinks they have a model problem. Usually they have an evaluation problem and have not called it that yet.

Foundation models changed what is possible on both sides of my work. A VLA model now gets for free what used to take a dedicated stack for every task, and an agent can read a company's own records and act on them without months of integration first.

What has not moved is the work around the model: getting data into a usable shape, keeping a cluster busy, building evaluation you trust enough to act on. That is where most programs stall, and it is rarely the part anyone budgets for.

So a lot of what I get hired for turns out to be data, infrastructure and deployment work, even when the brief said the model needed improving.

04 · Trajectory

Where I have been

  1. 2008

    Istanbul Technical University

    B.S. Aeronautical Engineering

    Top student of the university, on BP, Nippon Foundation and government scholarships.

  2. 2010

    UC San Diego

    M.S. Mechanical & Aerospace Engineering, AI & Robotics

    Powell Fellow. My first published work, on reinforcement learning for multi-agent systems, alongside distributed algorithms for multi-vehicle coordination.

  3. 2014

    Georgia Tech

    Ph.D. Aerospace Engineering, AI & Robotics · M.S. Mathematics

    AFOSR grantee. AI and robotics for autonomous aerial and ground vehicles, and estimation for multi-agent systems with unknown parameters. Finished all four degrees at 4.0.

  4. 2014–15

    NASA Jet Propulsion Laboratory

    Robotics Researcher

    Autonomy for unmanned ground vehicles running day and night over long periods, with AI-based perception, sensor fusion for localization, and coordinated path planning across groups of vehicles.

  5. 2015–17

    Mercedes-Benz R&D North America

    Senior Vehicle Intelligence Engineer

    Core self-driving technology for urban and highway scenarios, including AI-based prediction and the behavior architectures behind passenger safety and ride comfort.

  6. 2017–18

    Cyngn

    Head of AI, Perception, Sensor Fusion & Simulation

    Built and led the autonomous driving team, and brought the separate autonomy components together into one vehicle that worked.

  7. 2018–19

    XPeng

    Head of AI, Sensor Fusion, Simulation & Motion Planning

    Ran the AI and autonomy stack for electric cars and took driver assistance features into production, leading a large multinational team and building partnerships with US suppliers and universities.

  8. 2019–22

    Volax

    VP of Engineering

    Assembled a distributed engineering team for warehouse automation across computer vision, sensor fusion and real-time decision making. Added LSTM and transformer networks to trajectory prediction, and built the Spark and Airflow platform feeding continuous training.

  9. 2019–now

    Silicon Valley Consulting Group

    Founder

    Enterprise AI agents, AI factories and physical AI, built with an international team of engineers for manufacturing, logistics, energy, financial services and the public sector. Alongside it I am building PhysicalOS, an operating system for physical AI, and I also built Bldify, an agentic platform for field operations.

  10. 2022–now

    Deloitte

    VP, AI & Physical AI · Global CTO, AI Factory

    Built the AI Factory capability from GPU infrastructure up to the application layer, and shipped production agentic systems across several industries. Also technical CTO for autonomous vehicles and physical AI.

05 · Now

Working together

I take on a few engagements at a time. Usually a team has something working in a notebook or a demo and needs it to run properly, or they are deciding what infrastructure to commit to before they spend the money.

Consulting goes through Silicon Valley Consulting Group, which I founded and run with an international team of engineers. We build enterprise AI agents, AI factories and physical AI systems, mostly for manufacturing, logistics, energy, financial services and the public sector. I also founded Silicon Valley Investment Group, and I speak at conferences and seminars on physical AI and agentic systems.