The lab building Anant.
Neural AI studies how AI systems should remember, retrieve, update, and govern knowledge over time. Anant is the first product expression of that research.
Why the lab exists.
Modern AI systems are powerful inside a single interaction, but weak at continuity. They often treat work as disconnected prompts, even when the user is operating inside a long-running project, institution, or research process.
Our work focuses on persistent state: what should be stored, how it should be retrieved, when it should be updated, and how uncertainty should be represented when the system answers from memory.
The goal is not a larger chat history. The goal is a reliable memory layer for serious AI workflows.
What we work on.
Anant brings the lab's research into product form: structured memory, evidence-aware retrieval, consolidation, and deployment controls for sensitive environments.
Structured memory
Represent conversations, documents, decisions, and user context as retrievable state.
Evidence-aware recall
Attach recalled claims to sources, timestamps, and confidence signals instead of loose summaries.
Controlled deployment
Make the system work under real constraints: privacy, access control, audit, and data location.
How we operate.
Prototype in public surfaces
Research becomes visible through usable web, CLI, and institutional product surfaces.
Measure before scaling
We test recall quality, source faithfulness, latency, drift, and failure modes before expansion.
Design for governance
Every memory system needs deletion, inspection, permissioning, and operational boundaries.
Built from IIT Madras, For the World.
We are based at the Sudha and Shankar Innovation Hub, IIT Madras, Chennai. The lab was founded in 2025 to build original AI systems research from India.
Indian Institute of Technology Madras
Chennai 600036
India
From India
Built here, on purpose.
Neural AI is built in India, not as a marketing detail, but as an architectural commitment. The work we are doing — constructing cognitive systems that persist, reason, and remember — has reasons to be done from this specific place at this specific time.
01
Why this work should happen here
India produces a substantial fraction of the world's AI researchers and engineers. It has not yet produced a research lab that makes architectural-level contributions to how AI systems are built. The country has companies that deploy AI, integrate AI, and consume AI from foreign providers. What it has not had is a lab whose research output is read alongside Anthropic's, DeepMind's, or Mila's.
This is not a complaint. It is an observation about a gap that should not persist. The next decade of AI will be defined as much by architectural choices as by parameter counts, and there is no reason those architectural choices should be made exclusively in three or four cities outside India.
02
What Indian context requires
Most AI systems are built with implicit assumptions about the contexts they will operate in: monolingual or predominantly English, deployed via foreign cloud infrastructure, evaluated against English-language benchmarks, governed by data-protection regimes designed in other jurisdictions.
Indian cognitive contexts are different in ways that matter architecturally, not just linguistically. Code-switching between languages happens within sentences, not across documents. Names, relationships, and family structures follow conventions that do not map to English-language training data. Privacy and data-residency expectations are increasingly governed by Indian law. Building AI for these contexts is not a translation problem. It is a representation problem — and representation is an architectural decision, not a deployment detail.
03
What sovereignty actually means
Sovereignty in AI is often discussed as a policy concern: where data is stored, which jurisdiction governs it, who can access it under what circumstances. These are real questions and we take them seriously.
But there is a deeper sense of sovereignty that matters more for the work we do. A system whose cognitive architecture is built elsewhere, whose memory is shaped by training data we did not curate, and whose reasoning is opaque to us, is not sovereign in any meaningful sense. The deployment can be local. The dependency is not. Cognitive sovereignty — the ability to build, study, and modify the architectures that determine how AI represents the world — is the kind that compounds over time.
In practice
What this means in practice.
For Neural AI, building from India means three things concretely. First, the systems we build operate natively in Indian languages and code-mixed contexts, not as translation targets. Second, the architecture is studied and modifiable here — the substrate models we use are open-weight, the structured cognition we add is our own work, and the entire system is inspectable. Third, we build for deployment patterns that respect Indian data sovereignty by design rather than by retrofit.
These are not features added on top of a foreign architecture. They are consequences of where the architecture is built and by whom.
Scope
What we are not building.
A research lab is defined as much by what it refuses to build as by what it pursues. Several adjacent problems are being solved well by other organizations. Several others are not the kind of work this lab is structured to do. We list the major ones explicitly so that our scope is unambiguous to anyone who works with us, hires from us, or builds on us.
01
We are not training foundation models.
Frontier model training requires capital, infrastructure, and research depth that is concentrated in four or five organizations globally. The marginal contribution of a new entrant at that layer is small. Our work assumes foundation models exist and focuses on what should be built around them — the cognitive layer that current models lack.
02
We are not building agent frameworks.
Agent orchestration, tool calling, and multi-step task execution are being developed by capable teams across the industry. Our system can be invoked by agent frameworks or used alongside them, but we do not consider agent infrastructure our research domain.
03
We are not adding memory as a feature on top of existing chat interfaces.
ChatGPT and Claude have shipped memory features that serve consumer chat use cases. These are valid product additions for those platforms. Our work is at a different abstraction layer — institutional cognitive memory designed for organizations, not personal memory for individual chat users.
04
We are not building a cloud-only SaaS product.
The institutions that most need cognitive memory — regulated industries, governments, large enterprises with data residency requirements — cannot use cloud-hosted products that route their context through external infrastructure. Sovereign deployment is the architecture, not a deployment option we offer reluctantly.
05
We are not competing with enterprise search.
Glean and similar products do enterprise search well. Their unit of work is finding documents across an organization's tools. Our unit of work is structured cognition over the knowledge inside those documents and conversations. The two are complementary, not competing.
06
We are not building a general-purpose AI assistant.
Anant is not a chatbot, copilot, or virtual employee. It is the persistent cognitive layer that any AI assistant or human user can query for institutional context. The distinction matters because it determines what we optimize for: depth and durability of memory, not breadth of capability.
In sum
What this allows us to do.
Each refusal narrows the problem space we are responsible for. The narrower the scope, the deeper we can build into it. Our scope is exactly one thing: the cognitive memory architecture that organizations need to retain, structure, and reason over their institutional knowledge.
Everything in this lab is in service of that single problem.
People.
Neural AI is led by a small founding team and works with advisors, collaborators, and partners across research and deployment.
Work with the lab.
We collaborate with researchers, builders, institutions, and technical teams working on long-running AI workflows, governed deployments, evaluation, and knowledge-intensive systems.
Research
Partnerships
Evaluation