In this exclusive conversation with Rajeev Ranjan, Editor, Digital Terminal, Anuj Gupta, Founder & CEO of KiteFishAI, explores the shift towards smaller, specialised AI models and their growing relevance for enterprise applications. He discusses how sovereign AI can help Indian organisations strengthen data control, improve deployment flexibility, and address industry-specific needs across BFSI, healthcare, and pharma, while highlighting the key steps India must take to build globally competitive AI models and a stronger domestic AI ecosystem.
Rajeev: Why do you believe enterprise AI is moving from large general-purpose models towards smaller, specialised models?
Anuj: The shift is being driven by a practical reality: enterprises need AI that delivers business outcomes, not just impressive benchmark scores. Large general-purpose models have made remarkable progress and will continue to play an important role, particularly in complex reasoning, research, and tasks that require broad knowledge. However, deploying them across every enterprise workflow can be expensive, introduce unnecessary latency, and create challenges around data governance, infrastructure, and operational control.
Most enterprise problems are much more specific. A bank needs to understand financial documents, assess risk, detect anomalies, and interpret regulatory requirements. A hospital needs AI that understands clinical terminology, medical workflows, and patient records. These applications do not always require a model with hundreds of billions of parameters. They require a model that performs reliably on the task at hand.
Smaller, specialised models can be trained or adapted for particular domains, evaluated against relevant business requirements, and deployed closer to where the data resides. This can improve cost efficiency, reduce latency, and make deployment across private infrastructure more practical.
However, smaller does not automatically mean better. A poorly trained small model can perform worse than a general-purpose model. The real opportunity lies in building models with the right architecture, training data, post-training methods, and evaluation framework for the intended application.
At KiteFishAI, we believe the next phase of enterprise AI will be defined by specialised intelligence. Organisations will increasingly combine general-purpose models, domain-specific models, and smaller task-specific models rather than depend on a single model for every requirement. The question enterprises should ask is not how large a model is, but how effectively it solves a real problem at an acceptable cost and with the required level of control.
Rajeev: Where do SLMs offer a practical advantage for Indian enterprises dealing with sensitive or regulated data?
Anuj: For Indian enterprises, the advantages of small language models extend beyond inference costs. They create more options for deploying AI in environments where data protection, operational control, and infrastructure constraints are significant considerations.
Consider banking, insurance, healthcare, and pharmaceutical research. These sectors process financial records, personal information, patient histories, claims, research documents, and other sensitive data. Sending such information to an externally hosted AI service may introduce concerns around data residency, contractual obligations, confidentiality, and regulatory compliance.
SLMs can make it more practical to deploy AI within an enterprise data centre, a private cloud, or an appropriately secured on-premises environment. When inference happens within the organisation controlled infrastructure, enterprises can retain greater control over data access, logging, retention, and system integration.
There is also an economic advantage. A model optimised for a particular workflow may require fewer computational resources than a much larger general-purpose model. This can make it more feasible to deploy AI across multiple business units, branch locations, hospitals, or operational sites where network connectivity and infrastructure availability vary.
For example, an insurance company could use a specialised model to extract information from policy documents, support claims processing, and identify missing documentation. A hospital could use an internally deployed model to assist with clinical documentation or retrieve information from approved medical guidelines. In both cases, the model can operate within the organisation controlled environment, subject to appropriate access controls and human oversight.
That said, deploying a model locally does not automatically make a system secure or compliant. Security also depends on the surrounding architecture, access controls, encryption, auditability, data governance, and how the model is used. Healthcare and other high-stakes applications also require rigorous evaluation and appropriate human supervision.
For Indian enterprises, the practical value of SLMs is that they make secure and controlled AI deployment more accessible. They provide an additional option for organisations that need AI capabilities without making every workflow dependent on an external model API.
Rajeev: How can sovereign AI models help Indian organisations balance data control, model performance and deployment flexibility?
Anuj: Sovereign AI is often discussed primarily in terms of where data is stored or where computing infrastructure is located. Those are important considerations, but sovereignty has a broader meaning.
It includes the ability to understand, control, deploy, evaluate, and evolve AI models according to national and organisational requirements. It also includes access to the technical capabilities needed to develop and improve those models rather than relying entirely on systems controlled by external providers.
For Indian organisations, this matters because different applications have different requirements. A financial institution may need strict control over sensitive information and model access. A hospital network may need deployment across multiple facilities. A manufacturing company may need AI to operate within its private infrastructure, including environments with limited connectivity.
Sovereign models can support these requirements by giving organisations greater flexibility over where inference runs, how data is processed, which components are integrated, and how models are maintained.
Performance remains equally important. A sovereign model must not be adopted simply because it is developed domestically. It must demonstrate that it can perform the intended task at the required level of accuracy, reliability, latency, and cost.
This is where specialised model development becomes important. Models can be designed around relevant languages, domain-specific terminology, enterprise workflows, and local operating conditions. The objective is to build models that are competitive on measurable performance while providing deployment choices that meet organisational requirements.
At KiteFishAI, our approach is to develop foundational models from the ground up and build the capabilities needed to take them from pre-training through post-training, evaluation, inference optimisation, and enterprise deployment. We believe this approach creates opportunities to optimise models for specific requirements rather than treating a general-purpose model as the answer to every problem.
Sovereignty does not mean isolation from the global AI ecosystem. Indian organisations should be able to use open-source technologies, collaborate with international research communities, and integrate the best available tools. The important distinction is having meaningful control over critical capabilities and the freedom to choose the appropriate technology.
Ultimately, sovereign AI should give Indian organisations three things: control over their data and deployment, models that meet their performance requirements, and the flexibility to evolve their AI systems without being locked into a single provider.
Rajeev: What is the business case for domain-specific AI in sectors such as BFSI, healthcare and pharma?
Anuj: The business case for domain-specific AI starts with a simple observation: the value of AI is determined by the business problem it solves, not by the sophistication of the underlying model alone.
General-purpose models provide a strong starting point, but specialised applications often require domain knowledge, structured outputs, consistent terminology, integration with existing systems, and predictable performance. These requirements become particularly important in regulated industries.
In banking, financial services and insurance, domain-specific models can support document intelligence, customer service, policy interpretation, claims processing, compliance workflows, financial research, and risk analysis. A model developed for financial terminology and document structures can be evaluated against the actual documents and workflows used by an institution. This can reduce manual effort, accelerate processing, and improve consistency, provided the system meets the institution requirements for accuracy and oversight.
In healthcare, specialised AI can help with clinical documentation, medical information retrieval, patient communication, administrative workflows, and analysis of healthcare records. India has an especially diverse healthcare environment, with differences in language, infrastructure, patient volumes, and access to specialists. Efficient models that can be deployed securely across different settings could help make AI more accessible. Clinical decision-making, however, requires appropriate validation, professional oversight, and safeguards against incorrect outputs.
In pharmaceuticals, domain-specific AI can support scientific literature analysis, research document processing, pharmacovigilance workflows, regulatory documentation, and knowledge discovery across large collections of technical material. Models designed for scientific language and specialised information can help researchers and operational teams retrieve and process relevant information more efficiently. These systems must still be validated for their intended use, particularly where patient safety or regulatory submissions are involved.
Across these sectors, three factors determine whether domain-specific AI makes economic sense.
First, the model must deliver measurable improvements in task performance, turnaround time, or operational productivity. Second, the total cost of ownership must be competitive, including computing, integration, monitoring, maintenance, and governance. Third, the deployment model must satisfy the organisation security and regulatory requirements.
This is also why we should not evaluate AI economics using the cost of a single API call alone. Enterprises need to consider the complete cost of delivering a reliable outcome at scale.
At KiteFishAI, we see an opportunity to build specialised foundational models that can serve as a base for multiple enterprise applications. Instead of developing a separate model for every individual workflow, a strong domain-specific foundation can support several related use cases through appropriate post-training, evaluation, and integration.
The business opportunity is to move from generic AI experimentation to repeatable, measurable, and economically sustainable enterprise deployments.
Rajeev: What does India need to do to build globally competitive AI models and a stronger domestic AI innovation ecosystem?
Anuj: India has a significant opportunity to become a meaningful contributor to global AI research and model development. But achieving that ambition requires moving beyond AI adoption towards sustained investment in foundational research, infrastructure, talent, and original model development.
First, India needs sustained access to computing infrastructure. Training competitive models requires reliable access to GPUs, high-speed networking, storage, and efficient computing environments. Public compute programmes are an important step, particularly for startups and academic institutions that cannot independently fund large infrastructure investments. Access should be predictable enough to support experimentation, multiple training runs, evaluation, and iterative improvement rather than only isolated demonstrations.
Second, we need to invest in original research and model-building capabilities. Fine-tuning existing models is valuable and will remain an important part of the ecosystem. However, a globally competitive AI industry also needs teams that understand data curation, tokenisation, model architecture, pre-training, post-training, evaluation, and inference optimisation. India should encourage research into efficient architectures, multilingual intelligence, reasoning, embeddings, multimodal systems, and domain-specific foundational models.
Third, high-quality data must become a strategic priority. India has enormous diversity in languages, industries, scientific research, and real-world operating conditions. Yet useful data is often fragmented, difficult to access, inconsistently structured, or constrained by privacy and intellectual property considerations. We need mechanisms that enable responsible data access, high-quality dataset creation, benchmark development, and collaboration between industry, academia, and research institutions.
Fourth, universities and industry need stronger research partnerships. Universities can contribute fundamental research, specialised talent, and independent evaluation. Startups can translate research into working models and products. Enterprises can provide real-world problems, domain expertise, and deployment environments. Government programmes can help bridge the gap through research grants, shared infrastructure, and long-term collaborative programmes.
Fifth, we need to evaluate Indian models against global standards. Domestic development should not be treated as a substitute for performance. Models need transparent and reproducible evaluations across relevant benchmarks, languages, reasoning tasks, robustness, efficiency, and real-world applications. Publishing research, releasing datasets where appropriate, and contributing to open-source ecosystems can strengthen both credibility and scientific progress.
Finally, India needs a stronger commercial pathway from research to adoption. Building a model is only one part of the journey. Research teams also need access to enterprise customers, capital, deployment expertise, and opportunities to demonstrate measurable value. Procurement processes and innovation programmes should make it easier for credible domestic AI companies to run pilots and graduate into production deployments.
At KiteFishAI, we are pursuing this research-first approach by developing foundational models from the ground up, with an emphasis on specialised intelligence and efficient enterprise deployment. Our ambition is not simply to reproduce what already exists, but to contribute original research and build models that are competitive on their technical merits.
India does not need to replicate the exact strategies of the largest global AI laboratories. It can build strengths around efficient models, multilingual capabilities, domain-specific intelligence, and applications shaped by its own economic and social needs.
The long-term objective should be an ecosystem in which Indian researchers and companies can develop original AI technology, demonstrate it against global benchmarks, commercialise it successfully, and contribute to the advancement of AI worldwide. That is how India can move from being a major consumer of AI to becoming a meaningful builder of it.
𝐒𝐭𝐚𝐲 𝐢𝐧𝐟𝐨𝐫𝐦𝐞𝐝 𝐰𝐢𝐭𝐡 𝐨𝐮𝐫 𝐥𝐚𝐭𝐞𝐬𝐭 𝐮𝐩𝐝𝐚𝐭𝐞𝐬 𝐛𝐲 𝐣𝐨𝐢𝐧𝐢𝐧𝐠 𝐭𝐡𝐞 WhatsApp Channel now! 👈📲
𝑭𝒐𝒍𝒍𝒐𝒘 𝑶𝒖𝒓 𝑺𝒐𝒄𝒊𝒂𝒍 𝑴𝒆𝒅𝒊𝒂 𝑷𝒂𝒈𝒆𝐬 👉 Facebook, LinkedIn, Twitter, Instagram