How DeepLure Is Building AI That Actually Works In The Real World

Every few months, a major AI lab announces a new benchmark. A new model. A new score that edges closer to human performance. The press releases arrive, the valuations climb, and somewhere in a research lab, the previous generation of models is quietly deprecated.

Harsha Vardhan watched this cycle for years, first as an engineer at Wipro, then as a Master’s student at IIT Delhi and arrived at an uncomfortable conclusion: the world’s most well-funded AI companies are solving a problem that may not be solvable, while ignoring a category of problems that already are.

That gap is what DeepLure is being built to fill.

The Observation That Became A Company

Harsha Vardhan graduated from IIT Kharagpur in 2017 with a degree in Mechanical Engineering and joined Wipro’s CFO Office, where the Head of Finance pulled him into one of the earliest Robotic Process Automation initiatives in Indian corporate finance. It was 2017. Most businesses had no coherent AI strategy. Harsha was, almost by accident, building one.

Natural Language Processing sat at the core of that work. And the deeper he went, the more a specific discomfort took hold.

“All these NLP algorithms didn’t really comprehend language,” he says. “They were performing mere pattern matching.”

The systems were doing something statistically impressive and intellectually shallow and the people running them either didn’t see it or didn’t want to.

He filed the observation away and left Wipro in 2019, intending to pursue Human Centred Design in Denmark. COVID cancelled those plans. Left without a job or a programme, Harsha did something unusual: he spent nearly two years reading AI research from scratch, starting from the foundational papers of the 1950s through to 2020. Not structured courses. The original source material.

By the end of it, he had a clear direction “Neuro-Symbolic AI”, a field that attempts to combine neural networks with symbolic reasoning and a clear destination. The only faculty working on it in India was at IIT Delhi. He enrolled in August 2022, officially for a Master’s in Artificial Intelligence, but functionally to get inside the same rooms as the people working on the problem. Three months later, ChatGPT launched.

Betting Against The Hype

The response to ChatGPT reshaped the global technology landscape almost overnight. In India, AI-led investment accounted for 91% of all deep tech funding in 2025, with Indian tech startups collectively raising $9.1 billion,  a 23% increase year-on-year, per Nasscom-Zinnov. Globally, the scale was even more staggering. U.S. deep tech startups raised approximately $147 billion in 2025. The race toward Artificial General Intelligence had become the defining industrial project of the decade.

Harsha was unimpressed. Six months into the ChatGPT era, he had arrived at a view most of his peers were running hard in the opposite direction from.

“I realized that this problem cannot be solved,” he says. “It doesn’t matter how much we compute or how much data we throw at these algorithms, they will never truly comprehend language like us humans, since they lack human experience.”

What frustrated him as much as the technical reality was the institutional narrative around it. The largest AI labs in the world were advancing the claim that AGI was imminent. Either they were being deliberately misleading, Harsha thought, or they simply didn’t understand their own systems’ limits.

“Either all these intelligent engineers and scientists are lying, or they simply don’t know,” he says. “It’s actually the latter.”

The opportunity he saw in that gap was precise. Twenty-five years of applied AI research, work that had produced genuinely useful, commercially viable techniques was being systematically overlooked because it lacked the narrative excitement of AGI. That research wasn’t being discarded because it didn’t work. It was being discarded because it was old.

“AI may never reach human-level intelligence,” Harsha says. “However, it can do some tasks really well and those tasks can be leveraged to build great products.”

DeepLure was registered in April 2025. The mission: transform AI from corporate spectacle into real-world intelligence.

Starting With What Works

DeepLure is structured around three business verticals: an AI services practice, a products division, and a venture studio. The company is fully bootstrapped, employs 10 people drawn from IITs and NITs across India, and operates out of the Ratan Tata Innovation Hub in Amaravati, Andhra Pradesh, with cloud infrastructure supported by OVHCloud.

The services vertical, currently generating active revenue from multiple clients, funds the longer bets in the products division. It is a deliberate architecture: use what the market will pay for today to build what the market will need tomorrow.

Healthcare is where DeepLure has placed those longer bets, and the choice reflects a specific reading of where AI’s pattern-recognition strengths map onto India’s structural gaps. India has over 600 million active internet users and one of the world’s fastest-growing diagnostics markets, yet its healthcare infrastructure remains significantly under-digitalised, with fragmented patient records, inconsistent screening protocols, and limited access to specialist care outside major urban centres.

The first product, codenamed Sparsh, targets early childhood development screening for children aged 2 to 6, sold on a B2B basis to pre-schools, play-schools, and primary schools. Most existing tools in this space focus narrowly on cognitive development or screen only for ADHD and autism, typically through tablet-based assessments. Sparsh is designed to cover all five recognised developmental domains like  language acquisition, vision development, motor development, cognitive development, and social development. No product currently serves this full scope at the institutional level in India.

The second product, codenamed Arya, takes on a more structural problem: the absence of a unified patient journey layer across India’s healthcare ecosystem. Existing telemedicine platforms address point-of-care consultation but leave the coordination problem like fragmented records, no standardised referral pathways, no single interface connecting hospitals, labs, pharmacies and insurers. Arya is designed to be that interface.

“We don’t intend to sell patients anything,” Harsha says. “We want to guide patients through the healthcare ecosystem.”

The platform will be permanently free for patients. Revenue comes from the healthcare businesses such as hospitals, labs, pharmacies, insurance agencies that benefit from patient routing and platform integration.

Both products are in the pilot preparation phase, with hospital partnerships being developed through RTIH Amaravati, targeting a Q2 2026 launch.

The Technology Underneath

DeepLure’s technical architecture differs from the foundation model approach that dominates industry thinking. Rather than a single large general-purpose model, the company builds an ensemble of task-specific AI components,  each optimised for high performance within a defined domain and governed by a meta-system that delegates tasks based on the objective at hand. It is a more modular, more interpretable, and more cost-efficient approach for specialised applications.

The company is also running a separate R&D track that could have broader industry implications: a proprietary SDK and driver aimed at achieving Nvidia CUDA-equivalent performance from AMD GPUs on AI workloads. GPU infrastructure is the single largest cost component in modern AI deployment. Nvidia’s pricing has become a structural constraint for startups and enterprises operating in cost-sensitive markets. A viable AMD alternative would materially shift the economics of AI development — not just for DeepLure but for the industry broadly.

On the commercial side, the venture studio arm rounds out the business. Companies and individuals can approach DeepLure with an idea or prototype; the company conducts market research, develops and launches the product, and manages ongoing operations in exchange for an equity stake. One client is currently active in this arm.

What Comes Next

A fundraiser is planned, but not yet. Harsha has made a deliberate decision to enter investor conversations with pilot data rather than projections. The sequencing is consistent with how everything at DeepLure has been built, evidence first and narrative second.

The ambition, stated plainly, is large. Harsha wants DeepLure to be the company that made AI genuinely accessible, not by building the most impressive research system, but by designing products that work for everyday users, support regional Indian languages through proprietary audio pipelines, and embed AI capability into tasks where it produces measurable value for real people.

In an industry that has spent the better part of a decade building toward an intelligence it may never achieve, it is an unfashionably grounded position. But it is also, arguably, the more honest one.

“Transforming AI from corporate spectacle to real-world intelligence,” Harsha says, “that empowers everyone.”