Research questions

These are open questions, not finished work. I keep three lines of research going outside client engagements, and they are connected: each is about making AI systems that an organisation can trust, explain and run itself. As repositories, prototypes, papers and talks become public, they will be linked here under each question. Until then, read this page as a statement of what I am working on and why.

Ontology for enterprise AI

The question

How can an organisation's entities, relationships and rules be written down once, so that models, agents and people work from the same definitions?

Why it matters

Most enterprise AI failures I see are definitional. The model is asked about a "customer" or an "approved supplier" and the organisation has four meanings for each. Retrieval returns the wrong one; the agent acts on it; a human catches it late or not at all. An ontology is the layer that makes the system's view of the business explicit and checkable.

What I am working on

Lightweight ontologies that a client team can own: small enough to maintain, formal enough for an agent to use as a constraint, and tied to the systems of record rather than to a document. How to derive them from existing schemas and glossaries, and how to test a model's answers against them.

Deterministic orchestrators

The question

How far can a system be made predictable while the model inside it is not?

Why it matters

Enterprises do not deploy what they cannot predict. The useful behaviour of a language model is probabilistic; the workflow around it, the permissions, the review steps and the state transitions, need not be. The orchestrator is where determinism can be restored.

What I am working on

Orchestration patterns in which every state transition is explicit and testable, the model is called only inside bounded steps, and the system's behaviour on a given input can be reproduced. Where neural components can replace rules without losing that reproducibility, and where they cannot.

Native tokenizers and embedding systems for edge devices

The question

Can small, purpose-built tokenizers and embedding models run where the data is, without a round trip to a data centre?

Why it matters

A great deal of enterprise data cannot leave the device, the plant or the country. Retrieval and classification on the edge needs embeddings that are small, fast and tuned to the vocabulary of the domain, not a general-purpose model shrunk until it fits.

What I am working on

Domain-native tokenizers built from the organisation's own corpus, and compact embedding models trained against them, evaluated on retrieval quality per watt and per millisecond on commodity edge hardware.

Working with me on research

I take on a small number of research collaborations with universities, labs and client teams where the question above overlaps with a real deployment.

Email: hello@mayurpatil.ai