Rajesh Narasimha

The pitch fits on a business card: any camera, any fleet, real-time driver risk. Rajesh Narasimha has been saying some version of that sentence since 2017, when he and Soumitry Ray started Edgetensor out of the North Dallas suburbs. The pitch is boring, which is the point. Fleet safety is a boring problem worth a lot of money, and the founders who win boring problems tend to be the ones who already know how to make a computer vision algorithm run on a chip that costs eleven dollars.

Edgetensor sells software that turns a commodity dashcam into a driver-monitoring system. Not a box. Not a bundled hardware SKU. Software. Face detection, distraction detection, ADAS runtime models, event tagging, liability scoring, face and license plate blurring for the sensitive footage. It runs on the edge, meaning the truck itself, and it does not particularly care whose camera the fleet already bought. That is the whole trick.

The company says its platform processes seven million miles of driving data a month, quotes a three-million-dollar-per-hundred-vehicles liability savings figure, and claims a seventy-two percent improvement in driving behavior. Those numbers come from Edgetensor's marketing, so read them with the usual caveats. The point of quoting them here is that they are the numbers Narasimha is willing to hang above the fold on his own homepage, and a founder's chosen metrics tell you what he wants to be judged on.

Where the CEO came from

The résumé is unusually deep for an early-stage CEO. Narasimha holds a PhD in Electrical and Computer Engineering from Georgia Tech, completed in 2007. His Google Scholar page lists more than forty publications, seven hundred twenty-eight citations, an h-index of fifteen. The topics wander: joint classification and segmentation of whole-cell 3D images. Denoising algorithms for biological electron tomography. A patent on merging multiple exposures for high dynamic range imaging. This is not a founder who wandered into computer vision because the term was fashionable.

Before Edgetensor, he did tours at Texas Instruments and Samsung Research America, both places where making vision algorithms run on tightly constrained silicon is the day job. Between those, he was instrumental in setting up the US R&D operation for metaio, a German augmented reality company. Apple acquired metaio in 2015 and quietly folded its technology into what eventually became ARKit. Narasimha's LinkedIn does not brag about this. It does not really need to.

What Edgetensor actually does

The core product line is a set of camera-agnostic AI SDKs bundled into named modules: FleetGo for the general fleet management dashboard, FleetShield for safety and compliance, Cloud AI as the platform layer, a Gen AI feature set for intelligent analytics, and vertical bundles for waste management, school-bus safety, insurance, and automotive tier-one OEMs. The website reads like a menu because that is what fleet buyers want: a menu.

Underneath the menu is one technical bet. Inference happens on the edge device. That means video does not have to be shipped back to a cloud to be understood, which matters for three reasons - latency, bandwidth cost, and privacy - and one non-reason, which is that it sounds cool in a pitch deck. Narasimha's version of the bet is more specific: the model has to be small and fast enough to run on the hardware the fleet has already bought. If the fleet has to swap out its cameras, you have already lost the sale.

The 2023 launch of Edgetensor 2.0 added generative-AI features to the dashboard: a liability risk scorecard, accident avoidance flags, false alarm reduction, driver and road event tagging with real-time safety metrics, a single-pane-of-glass KPI dashboard, predictive analytics that compare fleets against historical baselines, and privacy features including automatic face blurring and license plate blurring. Integration paths include APIs, Docker, iFrame, and SSO. The list is the sales pitch; the interesting engineering is the small-model inference underneath it.

The camera-agnostic moat

The video telematics market is crowded. There are box vendors, incumbents with legacy fleets of installed hardware, and a middle layer of camera OEMs. Edgetensor's chosen position is upstream of all of them, or rather, orthogonal to them. Sell the AI as software. Let the hardware be whatever the hardware is. This is a bet that in the long run the model beats the box, because the box gets cheaper every year and the model gets smarter every quarter.

It is also a bet that reads naturally to a founder who spent a decade making vision algorithms run under real hardware constraints. If your instinct is to squeeze the model into the chip rather than requisition a bigger chip, camera-agnostic is not a marketing choice, it is a personality trait.

The privacy layer, before it was cool

Face blurring and license plate blurring were baked in from the platform launch, not bolted on after a customer complaint. Fleet operators watching a driver in the cab, and pedestrians outside the truck, both have obvious reasons to want their faces out of the training data. Insurers and regulators have obvious reasons to want the same. Edgetensor's approach is closer to a default setting than a compliance checkbox, which turns a friction point into a feature line item.

None of this is glamorous, and that is part of what makes it durable. The market for fleet-safety AI is measured in trucks and buses and school runs, not vibes. If the model works and the buyer trusts the data pipeline, the software renews. If it does not, the fleet goes back to spreadsheets and hopes.

The oddness of the range

The most human thing about Narasimha's public record is the range of topics his research career touches. Whole-cell 3D image segmentation. Electron tomography denoising. HDR image merging. Video analysis and image enhancement across a two-decade run of publications going back to 1996. Then, at some point, he decided the interesting problem was watching truck drivers on the freeway. The distance from a biological electron tomogram to a highway camera is longer than it looks on paper, and shorter than it looks in the abstract. Vision is vision. The rest is which pixels you point it at.

What the company looks like now

Edgetensor employs somewhere around twenty people, split between the Plano headquarters and an office in Bengaluru. The company raised a Series A round with reported activity around March 2021 and continues to iterate on the 2.0 platform. The customer base leans toward fleets and fleet-adjacent verticals - insurance, automotive OEMs and tier-ones, waste management, school buses. Testimonials on the website come from operators at HDfleet and Fleet Operate, both fleet-technology companies operating inside the same broader stack.

The founding team is small. Rajesh Narasimha as CEO; Soumitry Ray as CTO. Both are technical. Both have been in the computer vision world for long enough to have watched the field cycle through several fashions. The company they built is unfashionable in exactly the right ways: an inference layer, sold to logistics buyers, running quietly on hardware nobody notices.

What is next

Publicly available signals suggest Edgetensor is in the phase every Series A vertical-AI startup is in right now: pushing the gen-AI feature set into the dashboard, chasing insurance and automotive OEM partnerships where a single deal moves the revenue chart, and trying to make the camera-agnostic pitch land with fleets that have already bought a competitor's box. The specific next milestones are not public. The direction of travel is.

The interesting question for the next couple of years is whether the software-only, camera-agnostic thesis holds against video-telematics vendors who bundle hardware, or whether Edgetensor eventually has to ship its own reference hardware to control the stack end-to-end. Narasimha's whole career suggests he would rather not. The whole career also suggests he knows exactly how to if he has to.