A warehouse can be an expensive place to miscount. One box becomes a discrepancy; a row of boxes becomes an investigation; a label that makes perfect sense to a supplier becomes gibberish to the software waiting at the next loading bay. The work is physical, but the mistakes are often informational. Kushal Pillay has built his career in that awkward gap between what a camera sees and what a system understands.
Pillay is a robotics and computer-vision engineer and a founder of Powerhouse AI. The company, founded in 2021 and accepted into Y Combinator’s Winter 2022 batch, makes warehouse automation look less like a mechanical ballet and more like a familiar gesture: lift a phone and take a picture. From that image, its software is designed to count inventory, read labels and documents, check placement, and turn the visible mess of goods into structured data.
He has a tidy description for it: “Google Lens for warehouses.” It is the sort of line that saves a product demo several minutes. Photograph what is in front of you; let the software work out what it is, how many there are, and whether it belongs there. The camera is ordinary. The interpretation is where the company lives.
The useful part of a robot
Before the startup, Pillay worked closer to the machines themselves. His public founder biography describes computer-vision models that helped robots predict human movement, optimise their routes, and detect possible vandalism or threats. Other professional profiles place him in robotics research and development at Gaussian Robotics, with earlier engineering experience connected to Tetra Pak. His education followed the same line of inquiry: instrumentation and control at Vishwakarma Institute of Technology, then electrical engineering at the National University of Singapore.
This is the orthodox route toward building more sophisticated robots. Pillay took an interesting turn. Robotics had shown him not only what automation could do, but where it could become awkward. Hardware was expensive. A machine made for one task could be inflexible when the task changed. Smaller warehouses might need better eyes and cleaner records without needing a new metallic colleague.
The answer was to separate perception from the body. Put what Pillay calls “robot-grade tech” into the hands of human workers. A phone is already portable, understood, and flexible. Computer vision gives it a useful new occupation. The worker remains in the loop, while the software takes on visual repetition.
A co-founder, then a company
Powerhouse AI emerged from Entrepreneur First in Singapore, where Pillay met co-founder Ivo Verhaegh. The pairing joined two useful forms of impatience: Verhaegh wanted to start a business, while Pillay had seen a technical capability waiting for a more flexible delivery system. They worked toward warehouse automation that could be affordable and manageable, then founded the company in 2021.
Y Combinator followed in early 2022. The company’s YC profile sets out four practical areas: digitising workflows, counting inventory, processing labels beyond conventional barcodes, and reading logistics documents. These are not glamorous verbs. Count. Scan. Check. Read. Glamour has never reconciled a stocktake.
Consider the label. Small suppliers may send goods with their own labelling systems, while the receiving warehouse expects another. The traditional answer is physical relabelling, a triumph of adhesive over imagination. Pillay has explained how computer vision can instead read the original label and create a virtual record that travels with the item through different systems. The existing piece of paper stays put; the understanding changes behind it.
Small is a strategy
Pillay’s most consistent public advice is not about models. It is about scope. “Start small with clear objectives and scale gradually,” he told the EasyPost podcast. For an operator, the first question is not whether to “adopt AI,” a phrase broad enough to hide a thousand invoices. The question is which specific job should improve.
Maybe it is demand forecasting. Maybe delivery orders need to become text that a warehouse-management system can accept. Maybe workers spend hours handling third-party labels. Choose the friction first. Then choose the input and the expected output. Finally, pick the measure: did accuracy rise, did the task become faster, did errors fall?
One repetitive task + one clear objective + one measurable KPI. Earn the right to expand.
This sounds almost suspiciously sensible in a field addicted to sweeping futures. Yet warehouses punish vagueness quickly. Goods continue to arrive while a new tool is being introduced. Workers need to understand why it is there. Managers need a result they can defend. Software must fit around an existing operation whose peculiarities are features of reality, not bugs in a slide deck.
Pillay’s examples make the discipline concrete. A team could begin with a delivery order, ask software to transcribe its fields, and pass the result into the warehouse-management system. Another could focus on a recurring count or a supplier label that the current scanner cannot read. None of these projects requires a ceremonial declaration that the warehouse has entered the age of AI. It requires a before, an after, and somebody willing to compare them.
Accuracy and efficiency are two of the measures he returns to. If a new label workflow is quicker but creates more errors, the improvement is counterfeit. If it is accurate but requires so much supervision that it adds labour, the arithmetic will eventually object. A clear KPI is not merely a reporting device; it is protection against falling in love with the tool. Pillay’s method gives novelty a probationary period.
Pillay is frank about expectations. Buying software does not mean buying a robot, and it does not mean reducing human involvement to zero. His assistant metaphor is important because it gives both the technology and the worker a defined place. Let the software handle the dull, repeatable visual task. Let people manage exceptions, judgment, movement, and the untidy bits that refuse to become a benchmark.
Inputs, outputs, and tacos
There is little performance in Pillay’s explanations. Asked how much technical knowledge an operator needs, he reduces the problem to two things: know the input and know the output you expect. The system can take care of the deep technical details in between. This is not an argument against expertise; it is an expert’s argument for a usable interface.
He is similarly approachable about the beginning of a learning curve. Search for something small on YouTube, he suggests, start from a keyword you understand, and let the internet lead you onward. Operators uncertain about what to use can find him on LinkedIn, where he has publicly offered to talk through their situation. For someone trained to make machines perceive the world, he spends a notable amount of time making the machines less intimidating to people.
And because every founder profile deserves one fact that cannot fit in a pitch deck: his comfort food is Mexican, usually tacos. Even computer-vision engineers occasionally prefer a problem wrapped neatly and served by hand.
The camera is the wedge
Powerhouse AI’s larger aspiration is visible in the modesty of its starting point. A warehouse camera can become a scanner that recognises more than a barcode. A photograph can become evidence for a count, a placement check, or an audit. A document can become data without a worker typing every field. Each use is narrow; together they form an increasingly legible picture of an operation.
The approach also preserves optionality. A phone can move to the awkward corner where fixed equipment cannot. Software can learn a new label format without sending a machine back to the factory. A team can test one workflow before touching the rest. This flexibility is the lesson Pillay carried out of robotics and into company building.
It also changes who can try the technology. Pillay and Bergsager have spoken specifically about smaller and medium-sized warehouses, where manual and semi-automated processes often sit beside each other. Such operations may not have the budget, appetite, or floor space for full-scale robotics. They still have invoices, inventory, misplaced goods, and a shift that would rather not count the same pallet twice. A focused visual tool lets them automate a fragment without pretending the rest of the building has vanished.
This is a quiet kind of product ambition. The software does not need to become the warehouse. It can become the useful layer between a worker’s glance and a database’s demand for order. Every successful capture supplies a little more visibility: what arrived, where it went, whether the count matches, and which exception needs a human. The camera is a wedge because everyone knows how to use it. The structured record is the prize.
His story is therefore not a retreat from robots. It is a careful edit. Keep the perception, discard unnecessary machinery, and place the capability with the person already doing the work. Begin with the box in front of you. Take the picture. Measure what changes. The grand warehouse future can wait its turn in the queue.