Research at AksutraWe study how AI systems are grounded, evaluated and made dependable — and how visual realism is perceived. Our long-term aim is foundational AI. Today that means careful research, published openly, and systems we put into real use.
Independent · Owned end-to-endFocus areas
Our current research is concentrated in four areas.
Grounded language systems
How language models can be made to answer from evidence rather than guess — and to say so when they do not know. This work runs directly into the agents we deploy.
Generative systems
Research into generative systems for text, image, and creative work. Including grounded generation, anti-extractive formation, and originality-preserving methods.
Applied intelligence
Applied AI research for enterprise and public sector — including systems designed for accountability, auditability, and integration with existing workflows.
Evaluation and safety
Research into how AI systems are measured, evaluated, and held accountable. Including multi-metric evaluation, calibration, and methods that prevent capability claims from outpacing measured results.
Published workOur papers are published openly on Zenodo. Each can be read here, page by page, without leaving the site.
Hardik Zayne · February 2026
Introduces the Structural Depth Function, a measure of the internally coherent process layers recoverable from any physical surface or created artifact, and derives an asymmetry: the cost of fabricating many coherent layers grows exponentially, while the cost of detecting a missing one grows only logarithmically.
Hardik Zayne · February 2026
States that perceived visual realism follows from complete constraint satisfaction across physical, perceptual and informational domains — realism as a consequence of obeying the world, not as an achievement of detail. Published under CC BY 4.0.
Hardik Zayne · January 2026
Why a grainy silent film can feel real while a flawless render feels hollow: realism in cinema, photography, VFX and generative AI is a matter of honouring physical constraints, not of resolution or rendering power.
Approach
Independent
We build our own systems. Where a third-party model is the best tool, we use it — and everything around it, from evaluation to security, is ours.
Owned where it matters
Our systems, our evaluation and our clients' data stay under our control or the client's. Where Indian data residency or jurisdiction applies, deployments are designed accordingly — including on the client's own servers.
Measured
We evaluate our work across multiple metrics, not a single proxy. We do not make capability claims we have not measured.
Long-term
We ship research when the work is ready, not when the market demands a demo. Pace is deliberate.
Where we areOur long-term aim is foundational AI, and we are early on that path. We publish research when we can stand behind it, and we do not announce models, benchmarks or dates before the work is finished.
What we can share is the direction: AI research and products built in India, measured honestly, and accountable in their behaviour — with results that others can check.