Aniruddha Banerjee, co-founder and CEO of SwitchOn, sits down with Rohit Yadav to explain why factory floors got automated decades ago while quality inspection stayed manual, and what finally changed. He left Nvidia in San Francisco. His co-founder, who is also his first cousin, left a team building India’s moon rover. They started in a Bengaluru apartment, working on a problem their own family’s manufacturing business knew well.

SwitchOn builds AI for quality inspection, aiming at zero defect and zero customer complaint. The core product, Deep Inspect, combines cameras, lighting and compute at the line. Since commercialising in 2021 the company has grown roughly 3x year over year, and now works with 37 clients across about 80 sites and close to 300 deployments, with a team near 100 and around 40 percent of revenue from Europe, Southeast Asia, and North and South America. The system detects defects as small as 150 microns, runs at up to 1,200 parts per minute, holds false positives under 0.5 percent, and lets customers train and deploy their own models in 45 minutes.

Whether you build hardware, invest in physical AI, or run operations in a plant, this one is about the unglamorous layer where most of the value actually sits.Chapters:

00:00 Two cousins, a family manufacturing business, and the quality gap

02:05 Why computer vision and inspection took so long to meet

04:01 What SwitchOn builds, and what came after Deep Inspect

05:27 The numbers: 37 clients, 80 sites, 300 deployments

08:25 Where deployments actually run into trouble inside a plant

11:58 Training on good product instead of defect data

15:23 Why inference runs at the line and not in the cloud

18:29 What stops a pilot from scaling inside a large manufacturer

22:07 Advice for founders: stakeholders, and willingness to pay

24:19 Hardware versus software in physical AI

27:38 The mid-market question, and why they have not gone there yet

31:13 Going global: product fit, referrals, and earning trust

34:50 The moat: 200 million defect images and what compounds

40:54 Closing the loop, and what DeepFix does

43:58 Where the value sits in 2031🔑 Key insights you will walk away with:

➡️ Why rule-based systems needed tens of thousands of images and still only handled a narrow set of checks

➡️ How transformers cut training to under 200 images and 45 minutes, and why that unlocked scale

➡️ The decision to train on good product rather than defects, and what it saves a customer who cannot stop the line

➡️ Why plants in remote locations and confidential product lines made cloud inference a non starter

➡️ What actually goes wrong on deployment, starting with teams who disagree on what good looks like

➡️ Why a pilot dies if value does not land in weeks rather than months

➡️ The founder advice on assessing willingness to pay before building

➡️ Why the hardware versus software split matters less than revenue quality and pricing power

➡️ How 200 million real defect images compound into something synthetic data cannot replicate

➡️ The case for special purpose systems over general purpose humanoids on a factory lineLinks:

Aniruddha Banerjee: https://www.linkedin.com/in/anirudhabanerjee/

SwitchOn: https://switchon.io/

Rohit Yadav: https://www.linkedin.com/in/rohityadav23/

Newsletter: https://yadavrohit.substack.com/