Many people believe that big tech companies in India primarily hire for coding roles, but Microsoft Research India challenges this assumption. The lab is working on generative AI projects covering foundational technologies, industry applications, and solutions to societal challenges in India.

“A common thread that runs through our work is an open, collaborative style where we partner with academia, government, NGOs, and more and share our work through open publication and open-source software,” said Dr Venkat Padmanabhan, managing director, Microsoft Research India, in an exclusive interview with AIM.

Building the foundations for generative AI

Padmanabhan shared that Microsoft Research is currently pursuing projects in retrieval and energy-efficient AI deployment. One of them is ReFoRM, which aims to improve retrieval and RAG for chat, search, and recommendation applications, especially over private enterprise data.  

“ReFoRM’s model seeks to deliver step-function gains in retrieval accuracy over all existing retrieval models akin to what LLMs have delivered for reasoning, while keeping the cost low,” he said. 

By comparison, Google also integrates RAG into its Gemini AI and Vertex AI RAG Engine platforms, for enterprise customers, combining large language models with advanced retrieval capabilities.

Meanwhile, social media giant Meta uses RAG to improve search and recommendation systems through large-scale embedding models and retrieval pipelines. Last year, the company introduced RAFT (Retrieval-Augmented Fine-Tuning) approach which combines Retrieval-Augmented Generation (RAG) with supervised fine-tuning to improve domain-specific adaptation of language models.

Alongside ReFoRM, Microsoft Research India is exploring sustainability challenges through Greenferencing, which combines renewable energy with intelligent load distribution across micro data centres. Padmanabhan believes such initiatives can make AI deployments both energy-aware and cost-efficient.

Complementing these initiatives, Microsoft Research India has been developing foundational systems that enable the next generation of AI applications.

One example is the DiskANN project on approximate nearest neighbour search (ANNS), initiated in 2018. It enables generative AI tools to embed documents, images, and user queries as high-dimensional vectors and find the most relevant matches efficiently. 

“DiskANN represents the state-of-the-art ANNS algorithm, able to serve an index of trillions of vectors at high quality and a fraction of the cost possible otherwise,” Padmanabhan said. He shared that the technology has influenced industry solutions, including Cassandra and Pinecone.

But the tool is not alone. Meta’s FAISS has become the default open-source choice for similarity search, driving recommendation engines and vector databases across industries. 

On the other hand, Google ScaNN is built for efficiency, with a strong focus on speed and accuracy. Unlike FAISS, which supports both exact and approximate search, ScaNN is dedicated solely to approximate nearest neighbour search, making it well-suited for scenarios where speed is critical.

Padmanabhan shared that Microsoft Research India is also working on generative AI for software coding, with two research papers published in this area. The first, CodePlan, combines LLMs with formal methods for program analysis to handle tasks that require coordinated changes across large codebases. 

The second, Code Researcher, conducts deep research on large system codebases and commit histories to generate solutions such as crash-preventing patches for the Linux kernel. “The research underlying both these systems has been published openly, sparking further research and development in industry and academia,” he said.

Collaborating for national impact

Microsoft Research India is contributing to IndiaAI’s mission to skill 500,000 individuals by 2026, including students, educators, women entrepreneurs, and government officials. “Our research solutions have already been piloted in partnership with NGOs and various state governments. We hope that working with the IndiaAI Mission would enable us to roll these out nationwide,” Padmanabhan said.

Earlier this year, Microsoft partnered with Yotta Data Services to accelerate AI adoption in India by integrating Microsoft Azure AI services with Yotta’s Shakti Cloud platform. 

Yet, despite these initiatives, startups under the IndiaAI Mission have not launched an LLM built from scratch. Meanwhile, Google also announced support for IndiaAI startups at its recent Google I/O event in Bengaluru.

Besides the startups selected under the IndiaAI Mission, Google is also working with others, including CoRover, Glance, Entri, InVideo, Nykaa, Dashverse, and Toonsutra.

The tech giant reiterated its support for India’s AI startup ecosystem through programmes like the Google for Startups Accelerator (GFSA), which has supported over 230 Indian startups. 

On the other hand, Microsoft Research India has undertaken extensive benchmarking of LLMs across Indian languages through initiatives like MEGA, Pariksha, and Updesh. 

Pariksha involved workers from Karya, an ethical data company, marking the first time rural Indians were directly involved in AI evaluation pipelines. “Karya workers contributed to evaluation and benchmark creation, bringing a wider set of perspectives into AI development,” said Padmanabhan.

The real question is whether Microsoft is truly solving India’s problems or just keeping pace with other global players. Google has been building Indian language datasets through Project Vaani, Meta has released open-source LLMs like Llama 4, which supports 12 languages natively, including Hindi. It provides full language capabilities such as translation, understanding cultural nuances, and generating grammatically correct text for Hindi.

Meanwhile, Microsoft’s research has focused on on-ground deployment, such as teacher training, healthcare assistants, and rural workforce involvement in evaluation.

Padmanabhan said that Shiksha Copilot, piloted in 2024 with 1,000 government schoolteachers in Karnataka, is now scaling to 8,000 teachers in Karnataka and Telangana. 

In healthcare, expert-in-the-loop AI assistants operate across hospitals and rural health systems in four states, reaching over 2,500 patients and 3,000 community health workers with multilingual, expert-verified guidance.

“Our approach is to see how AI can work in partnership with humans, with AI helping humans scale and humans helping ensure the accuracy of AI-generated answers,” said Padmanabhan. The lab also launched A4I, a partnership with Microsoft Philanthropies and IIIT Bangalore, to build open-source AI tools as digital public goods.

Microsoft Research India’s work shows that AI research can move beyond the lab and into spaces where it solves pressing, real-world problems.

The post Peeking Into Microsoft Research India’s AI Experiments appeared first on Analytics India Magazine.