ozari+ · the AI inside Ozari
Grid intelligence that learns every shift and never operates
ozari+ reads thousands of live signals and turns them into a clear picture, a forecast and a specific suggestion. It learns from what happens next, improves in governed generations, and leaves every decision to your people.
Your people, plus.
ozari+ is the intelligence inside Ozari. It watches, forecasts, explains and proposes. It has no command field and no route to the field, so the decision always stays with a person.
What it does
Useful on a quiet shift and a busy one
Operating picture
Frequency, voltage band, feeder headroom, generation mix, protection and data quality, summarised all the time.
Forecasts with uncertainty
Load and DER forecasts with a 95% band, so operators see how sure it is, not just the estimate.
Anomaly detection
Spikes, drift and flatlined signals, such as an inverter that stops changing while the sun is up.
Asset health
A ranked watchlist for transformers, capacitor banks, DER and breaker duty, from real operation counts.
Root cause
Groups related alarms, points to the likely source, and keeps a history of past advice and what happened next.
Reasons in plain language
Explanations generated on your own server by default. No internet connection is needed.
Self-improvement
Self-improving. Never self-approving
ozari+ is built to get better with use, inside strict limits. It can learn and it can propose. Only people can put a new generation into service.
- Observe Every forecast is scored against what really happened. Every suggestion is logged with the decision and the outcome. Available now
- Adapt Models recalibrate to your network’s own history, so forecasts and baselines fit your feeders, not an average grid. Available now
- Propose ozari+ drafts improvements to its own models, thresholds and rules, as change requests your engineers can read. In development
- Prove Each candidate generation is replayed against recorded history, then runs in shadow beside the current one. In development
- Promote People approve the new generation. It is versioned, recorded on the tamper-evident ledger and can be rolled back. In development
No model changes itself in service. Each loop starts from what the last generation learned, and every promotion needs a person.
Governance by design
Advice, with its reasons, and never a command
ozari+ is built so that it cannot become a control path, even by mistake.
- No command field. It cannot create, change or approve a command.
- Isolated enclave. Designed to run in the OT DMZ, fed by a one-way data flow.
- Confidence and reasons. Every suggestion shows how sure it is and the facts behind it.
- Every step recorded. Suggestions, acceptances and dismissals are logged with a time and a name.
- Sovereign. On your own server by default. A self-hosted language model is optional. A remote model is used only with your consent, with redaction and a record of each call.
One recommendation, end to end
From forecast to outcome, with every step on the record
This is how one suggestion moves through Ozari. ozari+ forecasts and explains. The DERMS engine works out the options from the network model. Two people decide. The command path acts. Then ozari+ checks its own forecast against what happened.
- The maths does the physics. Options come from the network model, not from a language model.
- ozari+ ranks and explains. It shows its confidence and the facts behind the advice.
- People decide. A request and an approval, from two different people.
- Every outcome teaches. Each forecast is scored against the metered result.
Forecasting
See the uncertainty, not just the line
Forecasts use a damped trend model with a 95% band. The default engine is statistical and rule-based, so it is predictable, explainable and fast.
It is not a large language model. Language models are optional and only add narrative. The numbers always come from the network model and your own data.
How control works
ozari+ proposes. People and interlocks decide
A suggestion becomes an action only through the same path as every other command: two people, safety tags, the control lease and the audit.
Guidance
Designed around the AI-in-OT principles
In December 2025, Australia’s ACSC, the US CISA and seven partner agencies published four principles for using AI in operational technology. This is how ozari+ is designed against each one.
| Principle | How ozari+ is designed |
|---|---|
| Understand AI | Every output shows its confidence and its reasons. The default engine is statistical and rule-based, so its behaviour can be explained and tested. |
| Consider AI use in OT | ozari+ advises. It has no command field and no route to the field, by design. |
| Establish governance and assurance | Inputs, outputs and decisions are logged. Remote models are off by default. New model generations are designed to need human approval before they run. |
| Embed oversight and failsafes | Two-person approval, interlocks and a tamper-evident audit apply to every operator command. The control room runs normally without the AI. |
Questions
Straight answers about ozari+
Can advisory AI keep up with fast control?
It does not need to. Fast loops such as self-healing and volt/VAR control run as deterministic automation. People arm them in advance, zone by zone, with interlocks and the audit. ozari+, the AI-enabled advisor, is never inside the loop. It watches, explains and proposes.
What happens if ozari+, the AI-enabled advisor, is wrong?
Each suggestion shows a confidence score and its reasons, so people can judge it. Nothing moves unless a person approves it, or unless people armed the automation in advance. Every forecast is scored against what happened, so errors are visible.
Does ozari+, the AI-enabled advisor, need the internet or a cloud service?
No. It runs on your own server by default and works with no internet connection. A self-hosted language model is optional. A remote model is used only with your consent, with redaction and a record of each call.
Is ozari+ a large language model?
No. ozari+ is an AI-enabled advisor, and its default engine is statistical and rule-based, so its behaviour can be explained and tested. A language model is optional and only adds narrative. The numbers come from the network model and your own data.
Can ozari+, the AI-enabled advisor, change its own models?
Not in service. Today it recalibrates its forecasts to your network’s own history. The propose, prove and promote stages are in development, and each new generation will need a person to approve it.
Where ozari+ is heading
Built for the next decade of the grid
Plans, not promises. We build them with pilot and design partners, in this order.
In development
Governed self-improvement
The propose, prove and promote stages of the loop: improvements drafted by the AI, proven in shadow, approved by people.
Final stages
Solar and wind plants
ozari+ for plants as well as networks, on the Ozari generation and renewables module.
Roadmap
Intelligence at the edge
Small models at the substation or plant that keep watch and store data when links are down.
Roadmap
Federated learning
Learn from many networks without moving their raw data. What is learned travels; the data stays home.
30–50%
Shorter outages that AI-based fault detection can deliver, according to the IEA.
No. 1
Barrier to AI in energy: a lack of digital skills, in an IEA survey of energy companies.
4
Principles for AI in operational technology, published jointly by nine agencies including ASD’s ACSC and CISA.
Start safely
Begin with a read-only shadow pilot
Ozari runs beside your current systems, sees the same data and issues no commands. You measure the value with your own data before anything changes.