One data model, with every answer traceable.
Sutra reads the systems a utility already runs, joins them in one model, and returns answers that carry their evidence.
What it connects to
Sutra reads what your systems already produce and replaces none of them.
Tariffs, eligibility rules and obligations run as deterministic code. A language model never interprets them.
Since 1-Apr-2026, every AMI deployment must be interoperable under the CEA guideline. The guideline's roadmap is REST messaging with JSON for IEC 61968-9 nouns. Sutra reads the IS 15959 register set by OBIS code and the IEC 61968-9 meter-reading fragment over JSON.
India is publishing standard schemas for meter data and regulatory filings, with their own validation scripts. We are building to them. We have not published a validation report, so we do not claim conformance. When we do, it goes on the stack page with its date.
How it is built
Connect
Sutra checks meter and grid streams on arrival, and quarantines a failed reading with the reason. It never fills one in. Pull connectors read billing and registry systems on your schedule.
Model
One versioned model of the grid defines every term: assets, meters, readings, filings, consent. Tests keep it in step with the code.
Answer
Each question goes to the cheapest engine that can answer it verifiably. When the evidence is missing, Sutra says so.
Prove
Every figure is labelled measured, modelled or illustrative, and opens to its formula and inputs. Any computation can be replayed and compared byte for byte.
Improve
A model change ships only if it beats the current model on a frozen test. Field results your team confirms retrain the models on your network.
All five run together on hardware you control. See where it runs.
How the AI is kept in check
EBRM, our router, sends each question to a calculator, grid physics, the knowledge graph or document retrieval first. A language model is used only to phrase the answer.
- Sutra checks every number in generated text against a computation it performed, and flags any figure it cannot trace.
- Forecasting models run beside our physics baseline. We show both answers and any disagreement.
- A person must approve regulator letters, disclosure drafts and spend recommendations. The code has no way to skip that step.
Built and tested on synthetic and public data. No field pilot has used it yet.
Where it runs
The whole platform runs on infrastructure you control.
Self-hosted
The application, database, graph store, cache and policy engine run as containers on your infrastructure. No managed cloud service is needed.
Your data stays with you
Sutra analyses grid, meter and consumer data where it lives. No other utility's data shares your database.
One outbound call
Only the language model call leaves your network. Sutra strips personal information first. Redaction is on by default and fails closed.
Works with the model off
Switch the language model off and Sutra returns deterministic answers, labelled as such.
The embedding and forecasting models load from local files. Refreshing India's published datasets is optional, and without it Sutra shows the date of the last load.
Running the language model on your own hardware is roadmap. The product includes a stand-in that reports itself as one.
From 1-Apr-2027, the CEA cyber security regulations require sensitive power-sector data, including cloud and historical data, to be stored encrypted and kept in India. They cover every distribution licensee. The 50 MW threshold applies only to generation and storage. On your infrastructure in India with the model call off, Sutra meets that residency clause. With the call on, a redacted prompt leaves.
The draft National Electricity Policy 2026 also asks for power-sector data to stay in India and for domestic software in critical applications. It is still a draft.
Start with a paid pilot.
One zone, with a measure we agree on day one. If the result is not better, we refund the fee. After the pilot, the license is a flat fee per utility, whatever your meter count.