Breaking
DIMAAG acquires Akridata to scale Physical AI - May 2026 Vision Assist · Vision Command · Vision Copilot now shipping Founder Kumar Ganapathy: two prior exits worth ~$1.2B $15M Series A led by TeleSoft Partners, Accel & MFV Partners Rule of the house: your data never moves DIMAAG acquires Akridata to scale Physical AI - May 2026 Vision Assist · Vision Command · Vision Copilot now shipping Founder Kumar Ganapathy: two prior exits worth ~$1.2B $15M Series A led by TeleSoft Partners, Accel & MFV Partners Rule of the house: your data never moves
Company · Applied AI · Industrial Inspection

The Startup Teaching Cameras to Catch What Tired Human Eyes Miss on the Assembly Line

Akridata spent years fixing AI's boring, expensive data problem - then pointed the same tools at factory defects and got acquired by DIMAAG in 2026.

There is a version of artificial intelligence that never writes a poem or paints a picture. It stares at a metal bracket sliding past on a conveyor, ten thousand times an hour, and flags the one that is cracked. It is unglamorous, repetitive, and enormously valuable - and it is the version Akridata decided to build. The company, founded in 2018 and acquired by DIMAAG in May 2026, spent its life on a problem most of the industry preferred to skip: not the model, but the mountain of visual data underneath it.

Akridata's tagline reads "Smarter Visual Data, Better Human Decisions." That ordering is deliberate. The company was never selling a robot that replaces the inspector on the line. It was selling a way to make the inspector - and the data scientist behind them - faster, cheaper, and more sure of what they were looking at.

2018Founded
$20MTotal raised
~23Employees
2026Acquired by DIMAAG

01 / The ThesisThe boring problem nobody wanted

Most AI stories fixate on the model. Akridata's founders bet on the opposite idea: that the real bottleneck in getting AI into production is data - specifically, the exabytes of unlabeled images and video that pile up faster than any team can sort, label, or make sense of. Their early product, the Edge Data Platform, was pitched as the first infrastructure built for "data-centric AI," spanning edge, core, and cloud to deliver datasets that were curated, consistent, and actually relevant.

The follow-on tool, Data Explorer, let data-science teams search and curate large unlabeled visual datasets without labeling all of it by hand first. The unromantic pitch: stop paying to annotate data that will not change your model. Curate first, label less.

Real-world AI is all about data, and solving AI's data problem is what Akridata was formed to do.Kumar Ganapathy, CEO & Co-founder
Akridata platform preview
The pitch, in one frame. Akridata's own preview image for the platform - the tidy version of a very messy problem: turning oceans of raw visual data into something a model can learn from.

02 / The ProductsThree tools, one engine

By 2025, that data engine had become a product line pointed squarely at the factory floor. Where Data Explorer was a tool for data teams, the Vision suite was built for the people who actually run quality control. There are three, and the division of labor is clean.

Human-in-the-loop

Vision Assist

An AI-assisted inspection system that sits alongside a human inspector, flagging defects in real time and taking multimodal input - image, text, voice. It plugs into existing legacy lines rather than ripping them out.

Command center

Vision Command

A real-time control room for manufacturing: tracks defects across lines, monitors compliance, keeps audit logs, and tunes throughput. This is the layer a plant manager watches.

Build & train

Vision Copilot

A collaborative platform for data scientists to visualize, curate, label, and generate synthetic data, then train, test, and deploy vision models - with the data kept where it lives.

Vision Copilot dashboard
Where the models get made. The Vision Copilot workspace - filtering, curating, and labeling visual data in one place, so a team can spend its budget on the images that actually move accuracy.

The labeling-cost argument

Vision Copilot's whole economic case is that labeling is where computer-vision budgets go to die. By intelligently prioritizing which datasets are worth annotating, the platform is designed to help teams "avoid overspending on data labeling" and reach a working model in fewer training iterations. And when real examples are scarce - a specific crack on a specific weld shows up rarely - it can generate synthetic data, including for hard verticals like railways, to teach a model what failure looks like before it happens in the field.

03 / The RefusalYour data never moves

One design decision did more selling than any feature. Akridata built the platform so customer data is not moved or copied - it stays wherever the customer chooses, cloud or on-premise, and Akridata itself has no access to it. For manufacturers in medical devices, railways, automotive, and government, where regulatory audit logs and data control are not optional, "the AI comes to your data" was frequently the entire reason to sign.

Your data isn't moved or copied - it remains where you choose.Akridata product principle
Akridata visual inspection interface
Defects, on the record. The visual-inspection view - the part customers in regulated industries cared about most, because every flagged fault is traceable and logged.

04 / The FounderA third act, on purpose

Akridata's CEO and chairman, Kumar Ganapathy, did not need another company. A graduate of IIT Madras with a PhD from the University of Illinois at Urbana-Champaign, he had already founded VxTel, acquired by Intel for a reported ~$550M, and co-founded Virident, acquired by Western Digital's HGST for a reported ~$685M. He co-founded Akridata in 2018 with Vijay Karamcheti - a Virident co-founder - and Sanjay Pichaiah.

The pattern across his companies is infrastructure, not applications: telecom silicon, flash storage, and now the data plumbing under AI. Akridata fits. It is the least fashionable layer of the stack, and the one that tends to decide whether the fashionable layers actually work.

05 / The MoneySmall, focused, acquired

Akridata raised roughly $20M across its life. The headline round was a $15M Series A in October 2021, led by TeleSoft Partners, Accel, and MFV Partners, aimed at scaling the team and the Edge Data Platform. The company stayed lean - around 23 people at acquisition - which is a notable stat for a company selling into heavy industry.

2018-19
Seed · ~$5M
Oct 2021
Series A · $15M
May 2026

In May 2026, DIMAAG acquired Akridata to scale what it calls Physical AI - the marriage of AI software, edge computing, hardware, and energy infrastructure for the physical world. Akridata brought the vision systems, edge processing, and inspection workflows; DIMAAG brought the hardware and industrial footprint. The combined pitch spans med tech, food tech, supply chain and logistics, automotive, rail, government, and industrial inspection.

06 / The MarketWhere it fits

Visual inspection AI is a crowded, practical corner of the market. Akridata sat between two camps: the data-curation and MLOps tools that help teams build vision models, and the shop-floor inspection vendors that catch defects in production. Its edge was doing both - and refusing to move your data while it did.

How to place Akridata

DoesReal-time visual defect & anomaly detection, plus the data curation behind it.
BuyersManufacturing, automotive, railway, medical device, government, agriculture; plus ML/data-science teams.
EdgeData stays in place; curate-first labeling; synthetic data; runs at the edge for real-time speed.
AlternativesLanding AI, Instrumental, Elementary, Averroes; data tools like Scale AI, Voxel51, Roboflow.
ModelB2B enterprise SaaS, cloud or on-premise, licensed per platform and workflow.
Vision Command pilot interface
Mission control for the line. A Vision Command pilot view. The command-center layer is where a defect stops being a single flagged frame and becomes a number a manager can act on.

07 / The TakeawayWhat you can copy

Akridata never became a household name, and it did not try to. But the arc is a template a lot of applied-AI startups are quietly following, and there are a few moves worth stealing. Chase the boring layer: the data problem was less crowded and more defensible than the model. Sell to the inspector, not against them - human-plus-AI was an easier sale than replacement. Make privacy a feature, not a compliance chore; "your data never moves" closed regulated deals. And know your exit shape: a lean, technical team with real industrial traction is exactly what a hardware-heavy acquirer like DIMAAG wants to bolt vision onto.

The honest caveat: this playbook works when your buyers already feel the pain and control real data - factories, hospitals, rail operators. In consumer markets, or where the data is thin and the defects are cheap to miss, the same discipline can read as over-engineering. Akridata's bet paid off precisely because it picked customers for whom a missed defect is expensive and a moved dataset is a liability.