Dragonfly · Physical AI Platform

    Most operations can already see what happens on their floor. Almost none can act on it safely. This is the story of a platform built to close that gap - and of the stack that had to grow up with it.

    Physical AI
    Azure · Databricks · Terraform
    Built on Claude Code

    The industry problem

    Cameras everywhere, decisions nowhere

    Warehouses, retail floors, ports, plants and studios have spent a decade installing sensors and cameras. The footage is there. What is missing is the layer between seeing and doing: something that understands what it is looking at, decides what should happen next, and can be trusted to say so in front of an auditor.

    That gap is not one company's problem. It runs across every operation where physical events drive money - stock that walks, a pallet in the wrong bay, a safety breach nobody logged. The usual answer is another dashboard, and a person left to read it. The industry needs software that acts, under governance, without handing control away.

    Proof the model works

    What Dragonfly has already earned

    50+
    Enterprise installations live today
    3:1
    Average customer ROI
    97%
    Customer satisfaction
    2-4wk
    Typical deployment vs. 6-12 months traditional

    The story so far

    A prototype that outgrew itself

    The first version was a no-code build: a working client layer around Dragonfly's AI Vision engine - secured access, an embedded dashboard, subscriptions, an API. It is still live today, and after a year of continuous feature work it did its job, which was to prove the commercial model rather than carry the industry one.

    Demand then moved past what a prototype stack can hold: many tenants, real-time reasoning over operational data, and security posture that enterprise buyers audit rather than trust. So the platform is being rebuilt from the ground up as a multi-tenant Physical AI platform-as-a-service, engineered for agentic systems and high-security data governance, with major industry players backing what it is becoming - locally and abroad.

    Computer vision in action

    The evolution

    From no-code prototype to an AI-native stack

    v1 prototype
    Proved the model
    rebuilt on
    Production rebuild
    Shipped in days, not weeks

    This migration is the part most teams underestimate. Moving from a visual builder to Azure, Terraform and Databricks is not a port - it is a change in how data is modelled, how environments are created, how identity and tenancy are enforced, and how far you can trust a machine to act.

    We made that journey alongside the product. The same team that shipped the prototype now builds AI-native, with agents in the loop of our own work. That is the capability we bring to anyone sitting on a validated idea that has outgrown the thing it was proved on.

    What's changing

    Five shifts, one governed platform

    One platform, many tenants

    A single multi-tenant PaaS replaces a bespoke build per client - provisioned once, extended per customer.

    Agentic systems, not dashboards

    AI agents reason over live operational data and propose actions, rather than just visualising it for a person to act on.

    Governed by design

    Every recommended action still needs a human yes before it reaches the physical world - autonomy with a human hand on the switch.

    High-security data governance

    A dedicated, audited path - separate infrastructure, separate credentials - carries every action from proposal to approval.

    Connects to any surface

    Teams, a CRM, a sensor feed, a legacy data source - the agents plug into whatever surface a business already runs on. The power is in the agents and the connectivity, not another dashboard to check.

    How it works

    Detection to decision, always with a human in the loop

    Step 1

    Edge detection

    A detection surfaces on-device, at the edge - no cloud round-trip required.

    Step 2

    Agent reasons

    An AI agent interprets the signal and proposes a recommended action.

    Step 3

    Governed path

    The proposal passes through a governed, audited path before anyone sees it.

    Step 4

    Operator notified

    A human operator is notified, with the recommendation and the reasoning behind it.

    Step 5

    Approved & actioned

    Only a human approval releases the action back into the physical world.

    The next chapter

    Where the platform is headed

    This is not being proved in a lab. Dragonfly is being built and deployed into some of the most difficult industrial operations on the planet - sites where a wrong call costs real money, stops a line, or puts someone in hospital. That is the bar the platform is engineered against, and it is why the governed path to act is the hard part rather than the dashboard.

    The direction of travel is speed without loosening the grip. Detection to decision is moving towards sub-second, so an agent has read the event, reasoned over it and put a recommendation in front of the right person while the situation is still live - and a human still approves anything that reaches the physical world.

    Behind the build

    How it's being built

    Deployed on Azure, provisioned with Terraform

    A multi-tenant footprint on Azure, spun up region by region with Terraform - reproducible infrastructure instead of hand-built environments.

    Databricks at the core of the data layer

    Dragonfly leverages Databricks for its core data infrastructure, so industrial data lands once and in one language. That foundation is what lets agents be deployed to the real world quickly, with a governed path to act on it.

    Built on Claude Code

    The new platform is built the same way we build increasingly much of our own work - AI-native, from the first line.

    The client

    Meet Dragonfly

    Dragonfly

    Dragonfly is a Physical AI platform already live in 50+ enterprise installations across logistics, retail, sustainability and media - and a 2026 CES Picks Award winner. Client: anisoptera.io

    Further reading

    Your Factory Already Has Cameras. Now They Can Think.

    Why the layer that pays for itself today is the software, not the robot.

    Read the article

    Building something that has to work in the physical world?

    Let's talk about agentic systems you can actually trust - with a human in the loop.