stavan.io Contact

Profile

Stavan Dholakia

Seattle, Washington, United States of America

I build safe, autonomous physical AI for hard environments. Fifteen years taking robot fleets, perception systems, and cloud silicon from founding team to production at global scale. Now building subsea robots at Mefferdi.

01 / Track record

Pilots are cheap. Site twelve is where you learn what you actually built.

My career has been spent at the point where a prototype has to become a product. The demo works, the money is committed, and the system now has to run every day, in places that do not forgive mistakes. That is the part I am good at.

Multi-$B
Infrastructure and commerce impact of programs I led or materially shaped across cloud, AI capacity, and automation
1.3M+ vCPUs
Arm64 cloud platform I own across generations, from silicon bring-up to production fleet
175K+ robots
Fleet covered by the functional-safety program I owned, working alongside people
43 sites
Robotic packing rollout I drove across North America and Europe, up from 5 pilots
40+ teams
~200 people in hardware, software, supply chain, operations, and commercial functions across North America, EMEA, and APAC
$450M ARR
Run rate of just one Azure program in a much larger portfolio

MicrosoftCloud silicon, AI infrastructure, and robotics

Cloud siliconShipped Cobalt 100, Microsoft’s first in-house Arm server CPU, now in 32 Azure regions. Leading Cobalt 200.
AI infrastructureNVIDIA GB200 and GB300-class GPU infrastructure for frontier AI: productization, capacity, deployment, adoption, and commercial execution.
RoboticsRobotics expertise for early-stage robotics work.

AmazonRobots and vision working next to people

Robotics · safetyOwned IEC 61508 functional safety for the robot fleet working alongside people, including a safety vest program with 99.97% human detection.
Go · Just Walk OutFounding member of the vision sensor platform team. Flagship grocery store launched in 10 months; pick-tracking sensor fusion at 99.2%.
Robotics · packingPacking cells and robot fleet from pilot to production, and a 6-DoF tote-handling arm from concept to production with +38% throughput.

CumminsHorsepower before compute

My first engineering job, and still the one I talk about most fondly. Cummins is where I learned that a design decision does not stay on a screen. It ends up on a highway, pulling 80,000 pounds, with a driver trusting it to start every morning. Everything I have built since carries that habit.

TrucksAdvanced vehicle electronics for Kenworth and Peterbilt trucks, worked out side by side with the truck makers.
PowertrainThe ISX and ISB engine platforms: the heavy-duty and medium-duty diesels that are workhorses of North American trucking.
ControlsSpecified and integrated the electronic control systems that tie engine and vehicle together, through global OEM launches. Promoted to senior engineer along the way.

02 / Now

Everything above was built inside large companies. This is what I am building with it.

Mefferdi

Co-founder · technical lead

Subsea inspection is still slow, expensive, and dependent on ships and crews. Mefferdi is an ocean-intelligence company building vehicles to change that: inspection that costs less, stays down longer, and sees more. Built by a team across four countries.

mefferdi.com

Design targets

1,000 m
Operating depth
Up to 16 h
Missions in near-freezing water

Vehicles and subsystems are built and in testing. Autonomy is in development.

~$4.1BEstimated offshore AUV and ROV market, 2026. Market context only.Mordor Intelligence

03 / How I work

Four things I do well

04 / Garage

From the garage

No lab manager, no procurement queue, no one to hand the hard part to. Just a bench, a power supply, and machines that either work or do not.

What I am building

One closed loop, at home: a rover that perceives and maps, an arm that touches and manipulates, and edge compute that has to decide without the cloud.

It is a small-scale version of the problem every physical AI company runs into. Perception arrives late, wrong, or uncertain, and it feeds a machine that can break things. I am building and testing the safety layer that sits between the AI and the motors, on real hardware, where a bad command has a sound.

Rig 01 · Manipulation

B601 arm

A six-axis arm with a gripper that I brought up from a kit and drive with my own controller instead of a vendor black box.

200 Hz
Custom joint-level control loop
CAN bus
Seven motors, integrated and tuned by hand
48 V
Actuation bus on bench power

What it took

  • Brought every joint online over CAN, one motor at a time, before anything was allowed to move.
  • Chased an elbow that dropped with a thud down to the raw CAN traffic: the stock software was cutting torque once a second and letting gravity win.
  • Wrote gravity compensation so the arm holds itself up instead of fighting itself.

What it achieves

  • Soft grasp and lift under my own control code.
  • A manipulation and safety testbed I understand end to end, from CAN frame to fingertip.

Rig 02 · Autonomy

Yahboom rover

An omnidirectional rover that maps a room, finds its way around it, and shows exactly what happens when perception falls behind the world.

LiDAR
Mapping and localization
Mecanum
Omnidirectional drive for tight indoor courses
ROS
Sensor fusion, mapping, and navigation stack

What it took

  • Out of the box it skidded, mapped badly, and drove the wrong way. I found four root causes: a mirrored LiDAR scan, an IMU reporting zeros from a firmware mismatch, sensor fusion that never received IMU data, and a navigation config fighting the robot.
  • The test course is masking tape at half-meter marks, a tape measure, and a mug slid across the path.

What it achieves

  • Maps a living room and navigates it on its own, returning to its start point within about an inch.
  • Wheel odometry within 0.8% over two meters.
  • Runs the same path under controlled perception delays, showing how stale data erodes safety.

Setup and fixes, open source

05 / Research

Physical AI research, built on real hardware

Alongside the industry work, I run an independent research program on one question: what makes a robot safe to deploy? It starts with perception, because most failures begin with a machine that is confidently wrong about the world, and it is moving toward manipulation. Six projects, open code, and real hardware in the loop.

Full record: Google Scholar · ORCID · GitHub

06 / Recognition and service

07 / Contact

Let’s talk.

If you are building or funding machines that have to work in the real world, I would like to hear about it.

Write to me

stavan@mefferdi.com

Connect on LinkedIn