I've spent nine years building electric work vehicles. Now I'm putting Claude in them.

What the role asks for, and what I've done

The LX40 and what runs on it

Boson builds electric utility vehicles for people who work outdoors: orchards, ranches, municipalities, campuses, and specialty-vehicle OEMs who buy the LX40 as a rolling chassis. Engineering and manufacturing are in Chennai; US final assembly is in Madera, California. I run the company from Frisco, Texas, and spend a lot of my time at a ranch test site near Bonham where the autonomous trucks log their routes.

Autonomy stack

• ComputeNeousys NRU-220S (NVIDIA Jetson AGX Orin), ROS2
• PerceptionZED stereo camera, 3D LiDAR, GEODNET RTK GPS
• PipelineBEV sensor fusion: RANSAC ground plane, obstacle grid, corridor generation, tree-trunk centerline for orchard rows
• SafetyIndependent obstacle / E-stop watchdog on a separate RK3588 SBC with a Coral Edge TPU and a depth camera, so the safety path does not share silicon with the planner
• StandardISO 18497 (2024, four parts): 12-page qualification checklist for the LX40 autonomy mode
• RoutesNamed GeoJSON segments recorded by the trucks, stitched into master routes; boustrophedon mower paths generated inside recorded boundaries

The compute question

Our per-vehicle compute BOM was $3,000 to $4,500. I designed a non-NVIDIA alternative around an RK3588 host and a Hailo-8 M.2 accelerator at roughly $1,100 to $1,400, and traced the real lock-in points to two places: the ZED SDK's CUDA dependency and YOLO-World. We decided to stay on Jetson for the next few years, but the exercise is exactly the kind of cost, latency and on-prem trade-off that physical-AI startups hit when they try to move a model from a dev box to a machine.

Founded

2017, with Suri Bhupatiraju

Sites

Frisco · Chennai · Madera

Test sites

Bonham, TX ranch · Lamb County, TX feed-pusher deployment

An operator talks. The truck proposes a route. A human confirms.

I'm building an MCP server that connects Boson's autonomous trucks to Claude. The design decisions so far are the ones I'd push any physical-AI startup toward:

• Talk to the ROS2 stack, not a fleet API. The model needs to see the same state the planner sees.
• Claude picks and stitches from recorded, already-traversed routes. It does not synthesize new paths. Every segment it can choose has been driven by a Boson truck.
• Nothing executes without operator confirmation. The route is a proposal until a person on the ground says go.
• A maneuver library, not free-form control. Scouting, perimeter patrol, mowing, return-to-base, reverse, U-turn. When the vehicle hits something it can't resolve, a remote observer proposes the next maneuver.

The next layer is an in-vehicle voice interface on the NRU: a conversation with the operator plus a screen that shows the data the conversation is about. Operators can also work from a phone through our web application.

Alongside the vehicle work, I use Claude daily on the business itself: board decks, ISO checklists, competitive teardowns of Monarch, GUSS, Agtonomy and Bluewhite, tariff and customs analysis, and route-processing tooling. I know what it is like to be a small team trying to get real engineering work out of a model, and where that goes wrong.

The hard part of physical AI isn't the model. It's the safety case, the latency budget, the on-prem compute, and the person who has to trust the machine. I've lived all four.

Context-switching is the job I already have

Off the clock I do my own mechanical work, including an IMS bearing and clutch on a Porsche 996. I like knowing how things fail.

Software first, then hardware

At Google I was a software engineer on Android Auto, on the database layer and on the OEM program that got the platform into BMW, Ford and GM vehicles. That was my first close look at how automakers make engineering decisions, and it shaped how Boson sells to OEMs.

I grew up in Andhra Pradesh, studied engineering in Chennai, and moved to the US in 2007. I've lived in eight states since, nine years of it in California, and the last three and a half in Texas.

2017 – now

Co-founder & CEO, Boson Motors

Before

Software Engineer, Google (Android Auto)

Education

B.E., Chennai

Let's talk about physical AI

Happy to walk through the autonomy stack, the MCP server design, or how a hardware startup actually adopts a model. Currently in Frisco, Texas, and ready to relocate to San Francisco immediately; nine years in California already, with our assembly plant in Madera.