Responsible AI
Intelligence with boundaries. AI should help people decide, not silently make consequential decisions without accountability.
Six principles
Every AI-related feature we build or integrate is held to these.
Explainable where appropriate
Outputs come with enough context for a person to understand why, especially when they inform a decision about people or safety.
Permission-aware
Models can only use data the user and the use case are authorised to access.
Context-aware
Results are tied to the identity, place, asset and time they relate to.
Auditable
Inputs, versions and significant outputs are logged where the deployment implements it.
Privacy-conscious
Minimal, purpose-limited data; aggregate and non-identifying by default.
Human-supervised where necessary
A person reviews and owns consequential decisions.
Where AI is, and is not, today
We separate research from deployed functionality.
| Area | Status | Notes |
|---|---|---|
| Smart Solar Bench V1.0 Base: edge AI, computer vision, on-device analytics | Not deployed | Not installed in V1.0. They remain development or future-variant capabilities until installed and validated. |
| DoubleSlit research (AI/ML, IoT, robotics, edge) | Research | Each project page states its maturity and whether it runs in any product. See DoubleSlit. |
| YNTR models, inference and agentic systems | Research | Research and engineering under DoubleSlit; not claimed as deployed in a product until implemented and validated. See YNTR. |
| Analytics in ViSE and MTRX | To be confirmed | Metrics are rule-based and documented with source, method, refresh frequency and limitations. Any machine-learning step is declared separately. |
Where the intelligence runs
The mode is chosen per application, not assumed.
Cloud
Managed infrastructure for workloads that benefit from scale.
Edge
Close to where the data is generated, to reduce latency and data movement, once installed and validated on a device.
Private and customer-controlled
Inside the customer's environment, for sensitive or regulated data.
Hybrid
A mix, with a defined boundary for what leaves the site.
For deployments in India, design goals include data residency, local intelligence, Indian-language capability and low-connectivity operation. These are goals scoped per project and are not statements that every deployment already provides them.
How data is used for AI
- Personal data submitted through this website is not used to train AI models.
- Data from a customer deployment is not used to train models for other customers without written agreement.
- Training and evaluation use data with a stated lawful basis, minimised, and anonymised or aggregated where possible.
- Public-space sensing is designed for aggregate, non-identifying measures such as counts, unless a use case, lawful basis and public notice say otherwise.
- We do not use biometric identification, face recognition or emotion inference on people in public infrastructure by default; any exception needs explicit legal basis, consent, a privacy review and a separate written agreement.
Human oversight and accountability
An AI output may alert, recommend or help prioritise. For decisions that materially affect a person, such as access, assessment, employment or safety, a named human reviews the output and is accountable for the decision. People can ask how an output was produced and can ask for it to be reviewed.
We test models on the conditions they will meet (for example Indian languages, lighting, climate and connectivity) before integration and record limits and failure modes. Models are versioned and can be rolled back.
Features such as anomaly detection, predictive maintenance and simulation are described as designed capabilities. They are available in a deployment only when implemented, tested and stated in its documentation.
Transparency and feedback
- Maturity labels on every AI-related page: concept, in development, prototype, pilot deployed.
- Illustrations and renders are labelled as such; AI-assisted imagery is never presented as a photograph.
- Material changes to AI features in a deployment are communicated to the customer.
Concerns about an AI feature, or a suspected harmful or biased output: hello@ivirth.in, subject "Responsible AI". We review every report and publish material corrections. Last updated 12 October 2026.
Discuss an AI use case
Tell us the problem and the data you have. We will say what is proven, what is research and what governance it needs.