From Aerospace Puzzles to AI Solutions
Arun Subramaniyan started his professional journey at GE Research right after completing his Master's and PhD at Purdue University in the United States. His initial work focused on jet engines, aerospace materials, and degradation modeling. The team often faced complex problems that traditional physics alone could not solve.
"Physics gives you the known knowns. Physics-plus-statistics helps with the known unknowns," Subramaniyan explains. "But there was always a remaining 10-15% that was truly unknown. You couldn't predict it until it happened."
The Deep Learning Experiment
This persistent gap pushed the team toward early deep learning approaches in 2010. They adopted a data-driven strategy without predefined questions. "We just said: learn what 'normal' looks like from the data, and flag anything that deviates as an anomaly," he recalls. "And it worked. It was the only method that worked."
This innovation occurred a year before AlexNet's breakthrough revived neural networks' credibility. AlexNet demonstrated that models trained on massive datasets could outperform human-designed rules. Subramaniyan immediately recognized this inflection point. "That was the first time a system learned features without humans telling it what to look for. And it didn't just win – it beat previous approaches by 20%," he says.
An Unplanned AI Career Path
Growing up in Coimbatore, Puducherry, and Chennai, Subramaniyan specialized in aerospace engineering from his Bachelor's at Anna University's Madras Institute of Technology. He never planned to work in AI. In fact, during his Purdue days, a professor discouraged the field when an Indian friend sought AI opportunities. "Tell your friend to pick a subject that will get him a job. AI is not going to go anywhere," the professor advised.
Yet AI pursued Subramaniyan throughout his career. In 2019, he joined Amazon Web Services to develop high-performance computing capabilities beyond traditional cloud infrastructure. There, he engaged with the rise of transformers, neural networks for sequential data like text and speech.
Scaling AI Infrastructure
His team built AWS's first GPU clusters as companies like Anthropic pushed scalability limits. "When Anthropic showed up and said they needed 10,000 GPUs, we thought they were joking. They were three people," he laughs. "We initially gave them 100, and then a few hundred at a time. But they kept scaling, and we realised they have a model that can really scale, and we helped them build the largest LLM."
In 2022, Subramaniyan moved to Intel, continuing his cloud and AI work. Two years later, Intel spun off its enterprise AI software business into Articul8, selecting him to lead the new company.
The Articul8 Philosophy and Enterprise AI
Subramaniyan believes AI models are becoming commodities. Differentiation now hinges on applications, orchestration, validation, domain knowledge, and trust. Articul8's platform embodies this philosophy, enabling multiple specialized models to collaborate like experts in a room.
"If you're designing an aircraft, you need one model that understands aerodynamics, another for structures, another for propulsion, another for controls. There is no one-model-wins world," he emphasizes.
This approach proves critical in domains with data not available on the open internet. "Industrial, aerospace, oil & gas, B2B financial services – the data lives in closed systems," Subramaniyan notes. You cannot scrape your way into competence here.
Growth and Funding
Articul8's customers include Franklin Templeton, Hitachi Energy, AWS, and Intel. Recently, the company secured over half of a planned $70 million funding round at a $500 million valuation. This marks a five-fold increase from its Series A round in January 2024.
Envisioning India's Heritage AI Model
Beyond enterprise AI, Subramaniyan envisions a "heritage model" for India. He argues India possesses over 5,000 years of written cultural history across languages, yet digital representation remains uneven. For instance, Bhojpuri suffers massive underrepresentation.
Heritage models would anchor AI systems in India's cultural and linguistic depth, facilitating meaning transfer across regions, not just translation. "If I can anchor my model on cultural heritage writings and poems, the overall model would be at the same level of quality in any language, even if you don't have data for it," he explains.
This model would enable someone to speak in Tamil while another hears it in Bengali with identical nuance. Subramaniyan sees commercial logic too. Content in media, advertising, or education could be created once and personalized during distribution.
Cultural and Global Impact
He highlights the issue with Western models interpreting Indian classics like Tirukkural. "The analogies are all Western, and those analogies then acquire a life of their own," Subramaniyan observes. "Imagine the heritage of India actually feeding the value systems in the world, compared to us accepting the value systems coming from the outside."
Subramaniyan asserts he has excellent talent for this project and estimates it could be accomplished under $100 million. His journey from aerospace to AI leadership now fuels a vision to embed India's rich heritage into artificial intelligence, creating systems that reflect and propagate indigenous values globally.
