There is an old temptation in startup stories to tidy the route after arrival. The abandoned turns vanish. Coincidence is promoted to vision. Failure gets a tasteful coat of paint and reappears as strategy. Angam Parashar’s route to the supermarket shelf resists that treatment. It includes a publishing tool that earned money but would not scale, a shutdown, three months away from operations, a broad enterprise-AI phase and, eventually, a specific problem hiding under fluorescent shop lights.
The continuity was not the product. It was the people. Parashar, Ankit Narayan Singh and Muktabh Mayank kept choosing one another while changing nearly everything else. Today Parashar is the Seattle-based co-founder and CEO of ParallelDots. Singh is its chief technology officer; Mayank is chief data scientist. Their flagship, ShelfWatch, analyzes photographs and videos of retail shelves so consumer-goods companies can see what is in stock, how much space a brand occupies, whether prices and promotions are correct, and whether a store has followed the plan.
This sounds like a neat destination. It was not an obvious departure point.
Before the shelf, there were samosas
Parashar studied mechanical engineering at the Indian Institute of Technology Kharagpur from 2007 to 2012, completing B.Tech and M.Tech degrees. He was already orbiting entrepreneurship. During university he worked with Ecozen Solutions, then a young energy company, and spent time around IIT Kharagpur’s entrepreneurship community. His campus life was not confined to business plans: badminton, squash, the mechanical engineers’ society and Formula SAE all appear in his public record.
Singh was an old college friend. They had lived in the same hostel wing and belonged to the same department. Mayank arrived later, through Parashar’s first post-college job at the analytics consultancy Opera Solutions in Noida. The three friends wanted to build something, though “something” had a generous definition. In 2012 their conversations wandered from selling samosas in Bengaluru to artificial intelligence. The samosas lost.
The first real idea came from a mundane frustration. Singh, returning from Australia and out of touch with Indian politics, spent hours following links to catch up. The friends imagined software that could assemble a chronology automatically. ParallelDots would use natural-language processing to pull relevant stories from a publisher’s archive and arrange them into an explorable timeline.
Even the name carried a joke about uncertainty. After days of unproductive naming, Parashar’s brother Deepak suggested “ParallelDots”: the founders spoke grandly about connecting dots while appearing unable to connect their own. The domain was available. In startup nomenclature, this counts as providence.
The courage to close a working thing
The archive product was not a fantasy. It had customers and revenue. Its problem was less theatrical: it was not scaling. In 2016 the founders shut the operation. For three months they mixed time off with brainstorming and returned to the drawing board, an object startup mythology often invokes but few companies are willing to sit before without a marker in hand.
They incorporated a new company in 2017 under the same name and initially went wide. ParallelDots offered enterprise AI services and considered a horizontal platform that could serve many industries. There were applications in text analysis, sentiment and intent, and projects across sectors. The founders knew breadth could become a refuge from choosing. They gave themselves a limited window to narrow the business.
Consumer brands made the choice less abstract. Their requests repeatedly led into physical stores, where execution was important and visibility was poor. An ecommerce manager can watch clicks accumulate by the minute. A consumer brand may still learn what happened on a shelf through a field representative’s visit, a set of photographs and a slow manual audit. The store contained plenty of information; the difficulty was turning it into data.
ShelfWatch became the focused answer. A photograph could reveal out-of-stock products, share of shelf, brand visibility, planogram compliance, prices and promotions. Instead of waiting for a report to travel up an organization, teams could inspect results closer to the visit. ParallelDots was no longer trying to make AI useful everywhere. It was learning one visual environment in depth.
Three people, complementary inconveniences
Parashar has consistently described the founders by their differences. He took the commercial role. Mayank, a machine-learning specialist, led the core science. Singh, a self-taught programmer whose career moved from mining at Rio Tinto to what Parashar teasingly called “data mining,” bridged technology and deployment. A complementary team is pleasant in a pitch deck. In practice, it means colleagues who notice different problems and occasionally prevent one another from enjoying an elegant mistake.
That arrangement helps explain why the partnership survived the original product. Parashar’s advice to founders has been unusually plain: do not start a company merely for the performance of starting one. The people around the idea must believe in it, and family and team matter when events become uncooperative. His more recent formulation is equally direct: “A complementary founding team is greater than the sum of its individuals.”
In May 2022, ParallelDots completed a $4.5 million Series A led by Btomorrow Ventures, with existing investor Multipoint Capital participating. Btomorrow later announced another $1.5 million round in 2023, including $1 million of its own capital. The money was assigned to product development, hiring and international expansion. These are ordinary uses of venture capital; reaching them after closing the first version of the company is less ordinary.
Accuracy is technical. Trust is social.
Parashar talks about ParallelDots as an applied research group with a deliberately vertical focus. Retail shelves are hostile little stages for computer vision: packaging changes, products overlap, lighting varies, photographs blur and stores improvise. The company says its foundational models are trained on large volumes of shelf imagery and can reduce model-training time from weeks to hours. Its public company history also records more than 14 research papers and patent filings. Parashar himself is named with Singh and Mayank on a pending United States patent application for a machine-learning self-billing system, published in 2024.
Yet his more revealing competitive claim is not about a model. It is “customer empathy.” Enterprise buyers are not required to become AI researchers before they can use software. A vendor must explain what a system sees, be transparent about deployment and fit the technology into work that already has deadlines, budgets and tired humans attached. “Customer empathy is an incredibly important quality when you are building AI software for businesses,” he has said.
This is the operator’s correction to the engineer’s seduction. Accuracy matters. So do integration, speed, legibility and the ability to turn a flag on a dashboard into a corrected shelf before the promotion ends. A clever model marooned from the workflow is merely an expensive opinion.
What comes after the photograph
Parashar’s current horizon stretches beyond a single still image. He has discussed augmented-reality store audits that recognize products as a worker scans a shelf, video analysis for operational events, and generative AI that can help people interrogate large retail datasets in ordinary language. He has called generative AI a “Swiss army knife for brands,” though the company’s history suggests it will be most convincing when used for a tightly defined job.
The ambition still has a pleasingly physical test. A model may be trained in the cloud, but its verdict must survive a crowded aisle, a dented packet and a shopkeeper who has placed tomorrow’s promotion beneath yesterday’s price. Retail is full of these small rebellions against tidy data. The attraction for Parashar is precisely that the work does not end with recognizing an object. The recognition has to become a useful instruction for a salesperson, merchandiser or manager. The distance from pixel to decision is where a demonstration becomes an operating system, and where customer empathy stops sounding like a slogan.
In 2025, Parashar said ShelfWatch had been deployed in 40 countries and the company had 120 full-time employees globally. ParallelDots lists Seattle as its headquarters, with operations in Gurugram and Prague. That geography mirrors Parashar’s own professional arc: educated in Kharagpur, tempered in Noida’s analytics world, building a company across India, the United States and customer markets farther afield.
The tempting conclusion is that the three founders finally connected the dots. Their better lesson is almost the reverse. They learned which dots did not need connecting. A news archive did not have to lead elegantly to a retail shelf. A mechanical engineering degree did not require a mechanical engineering career. A company could close without dissolving the partnership that created it.
Parashar kept the name, kept the co-founders and changed the question. The shelf answered.