What we are currently looking into.

A voice estate that only works in clean English is a pilot, not a deployment. Mid-sentence code-switching, regional accent, and a bad line are the normal case on our deployments, not the edge case, and an agent that degrades on any of them fails in front of a paying customer.

Building our own evaluation set from the calls our voice deployments actually handle – code-switched, accented, on ordinary telephony – and holding the Vakyam stack against it. Turn latency is a first-class requirement in that set, not something measured after accuracy.

What we do not know yet

• Whether one multilingual model beats routing to per-language models once language detection is counted in the latency budget.
• How far accuracy degrades on real telephony against benchmark audio.
• What a fair public scorecard for an Indic voice agent would even contain.

Enterprises are being asked to trust multi-agent systems whose failure modes are documented and unsolved: specification and system-design faults, agents talking past each other, and runs that never properly verify or terminate. Nobody signs off on a system they cannot examine after the fact.

Pushing the evidence trail already behind the agent platform toward something a reviewer can audit: what each hop received, what it called, what it returned, and where a person intervened.

What we do not know yet

• What a run-level pass or fail criterion looks like when no single step failed.
• How to catch a false consensus spreading between agents before it reaches an action.
• Whether replaying an old run against a changed model is a meaningful regression test.

MCP made tool access portable. It did not make it safe. Provenance of tool results, permissioning, and containment are exactly what an enterprise security review asks about before a single write is allowed.

Permission models and provenance on our own MCP connections, with execution kept inside the customer boundary rather than brokered through us.

What we do not know yet

• How to attribute an action across a chain of agents that each handled part of it.
• What a least-privilege default looks like for a tool the agent selected itself.
• Whether a tool result should be trusted differently from a user instruction.

Our vision and telemetry work happens on plants, sites, and cameras where bandwidth, latency, and data-residency rules all argue against the cloud. Our customers keep asking for the model to sit inside their boundary, so where that boundary limits us is our problem to solve.

Sizing what Arjuna, Yantra, and Vahana workloads actually need once the model has to sit on-premise, and where a smaller specialised model beats a larger general one.

What we do not know yet

• Where the accuracy-per-watt line falls for inspection tasks.
• Whether physics-informed models beat larger generic ones on sparse plant data.
• How to update a fleet of edge models without a maintenance window.