Skip to content
MicrosystemsPrecision engineeringP9e Microsystems
Shared platform · Data & intelligence

AI Platform

The AI platform carries the group's model training, deployment and monitoring, including inference at the edge where bandwidth or latency requires it. It also carries a rule the group takes seriously: AI is used where it earns its place, and deterministic methods are preferred wherever a function has to be verified or made safe.

How determinism shows up here

Assurance-critical functions use deterministic implementations. Where a model informs a critical action, a human stays in the loop by design.

What it does

Inside the platform

01

Training & experiment tracking

Reproducible training runs with data and model versioning, so a result can be regenerated rather than remembered.

02

Inference serving

Model serving in the cloud and at the edge, sized against the compute actually available on the device.

03

Model monitoring

Drift and performance monitoring in production, because a model that was right at deployment does not stay right.

04

The verification boundary

Where a function must be assured, the deterministic implementation wins. AI advises; it does not hold the safety case on its own.

How it fits together

The shape of the work

Where a model may decide, and where it may not

This is a boundary the group draws before a project starts, not after an incident. A model is excellent at telling you where to look and poor at being re-run tomorrow with the same answer — so it advises, and something verifiable acts.

A model may decide thisBeing right most of the time is genuinely useful here.Crop stress and disease classificationYield and irrigation forecastingDefect candidates on a vision lineAnomaly triage across a fleetGeospatial feature extractionRanking what an operator should look at firstA deterministic implementation holds thisBeing right most of the time is a failure condition here.Flight control and attitude loopsSafe-mode entry and interlocksProtection trips and thresholdsDosing, curtailment and actuation commandsAuthorisation decisionsAnything that has to be re-run and matchBOUNDARYCrossing the line is allowed in exactly one directionA model may propose an action on the right-hand side. A deterministic check or a named human then approves it. The model never holds the safety case alone.
Who builds on it

Used across the group

  • Vyovrinda geospatial analytics
  • Sasyaveda crop models
  • Industrial Automation machine vision
  • Savithru fleet analytics

Ownership

A shared platform. It has no natural home in any single unit, so the group funds it centrally and the shared engineering bench maintains it. Every company and division draws on the same version.

How we decide what is shared
Related

Other platforms

Firmware SDK

The common embedded foundation — hardware abstraction, drivers, secure boot and update, with timing you can reason about.

Read more

Device Management

Fleet identity, provisioning and health — knowing what is deployed, where, running what.

Read more

OTA Platform

Staged, reversible firmware campaigns across a fleet — the return path designed at the same time as the forward one.

Read more

Building something that needs AI Platform?

Tell us the constraints and we will come back with how this platform would carry them — or say plainly if it would not.

Start a conversation