The foundation

One foundation. Every decision on it is auditable.

Silos in. One semantic model. Coherent decisions out.

silos in · one semantic model · decisions out — with a refusal when the data can't carry the answer

Interoperable by design

Speaks to what you already run. Replaces nothing.

The pain is never missing data. It is five systems that each hold a piece of the truth and refuse to talk. Sutra reads what they already produce and replaces none of them.

Everything we speak, in one place

systems billing · cis ami · hes · mdm gis · network scada secure http · sftp onramps

standards dlms / cosem obis registers is 16444 meters

identity ies-aligned ids · consumer · asset · transaction

consent depa-pattern artefacts · purpose-limited · time-bound · revocable dpdp erasure & export

access role-based, per surface every read audit-logged

And the rules — tariffs, eligibility, obligations — execute as deterministic code, never as a language model's interpretation.

The AI, honestly

The language model answers last. Everything before it is checkable.

Most questions a utility asks are not language problems. EBRM, our energy-based router, sends each one to the cheapest engine that can answer it verifiably: a calculator, grid physics, the knowledge graph, document retrieval. A language model is reached only when phrasing is actually needed. And refusal is a route of its own: when the evidence cannot carry the answer, the system says so instead of guessing.

The model phrases. It never invents.

Every number in generated text is mechanically checked back to a computation the platform actually performed, with the formula and inputs on record. A figure that cannot be traced is flagged on the output, not smoothed over. This is enforced in code, not asked of the model politely.

Other models ride as second opinions, never silent swaps.

Time-series foundation models forecast load and flag anomalies beside our physics-grounded baseline, and both answers are shown with any disagreement stated. A model upgrade must beat the incumbent on a frozen evaluation before it ships, and it arrives side by side, never as a quiet replacement.

A human signs everything consequential.

Regulator letters, disclosure drafts and spend recommendations always require human review before they leave the building. That requirement is frozen in the type system, not a toggle someone can flip under deadline pressure.

Where this stands

All of this is built, running and tested in the shipped product, on synthetic and public data. No field pilot has exercised it yet. The first pilot is the shakedown, and we would rather tell you that here than have you find it out.

Where it runs

Your cloud, or entirely inside your boundary.

For sovereignty or connectivity constraints, the platform runs entirely within the DISCOM's own environment — including on-device.

on-device

Deeper

The questions a buyer actually asks.

How does data get in, and stay safe?

Through the standard export formats your systems already produce — over secure channels, role-gated once it lands. Where consumer data is involved, access rides on DEPA-pattern consent artefacts: purpose-limited, time-bound, revocable — with DPDP-aligned erasure and export built in. Integration specifics under NDA.

What does "shows its work" actually mean?

Every answer carries where it came from, how it was computed, and whether it can be reproduced. When the evidence isn't there, the system refuses instead of inventing a number.

What runs on public data vs. my own feed?

Regulatory Benchmarking and Procurement & Markets run today on India's published data. Pulse, Atlas, and Proof activate on your own feeds — and each surface says plainly which state it's in.

What do the agents actually do?

Standing agents run fixed, replayable checks — transformer stress, billing anomalies, compliance gaps — and bring what they find to a human with the evidence attached. They never act on the grid, and every run can be re-executed to the same result.

How we engage

Start with a paid pilot. If the after isn't better than the before, we refund you.

One bounded zone, forward-deployed, an exit clause from day one. The license that follows is flat per DISCOM — it does not grow as you add meters.

Book a pilot How we engage