ai workflow

Field data in.
Remote HMI out.
AI takes an hour.

The site survey and data collection take half a day to a full day on their own. The frames and tag map the Falcon captured, plus screenshots and notes on the existing HMI, go to an AI agent — Claude, Codex, Antigravity and others — which produces a remote HMI program connecting to the Falcon over TCP or MQTT within the hour. Once it's presented and we understand the customer's direction, AI builds the dashboard and SCADA.

pipeline

timeline

Two stages: on-site prep, then AI.

The site survey and data collection take half a day to a full day on their own; AI turns the collected data into a remote HMI in 60 minutes — not including the dashboard / SCADA, which AI builds next once the customer's direction is understood.

PHASE 1 · ON-SITE PREP Half a day – a day

— Collect machine logs while it runs
— Photograph the HMI while it runs

Hand the collected data to the AI agent

PHASE 2 · AI PROCESSING 60 minutes

00 – 15 MIN

Match the snapshots to find the tags in the log

15 – 40 MIN

Build the remote HMI from the snapshots

40 – 60 MIN

Replay the log to verify and tune details

Output: a remote HMI program verified by replaying the log — not the dashboard / SCADA.

what's next Once the remote HMI is presented and we understand the customer's direction, AI builds the dashboard / SCADA.

why it works

This job happens to be an AI sweet spot.

The input is structured

The Falcon does not hand over vague requirements — it hands over concrete frames, addresses and value ranges. That gives the agent real evidence for inferring what each register means.

The screen is the spec

Field labels, units, buttons and indicator lamps on the existing HMI amount to a ready-made UI specification and naming convention, which maps straight onto tags and layout.

Results verify instantly

Connect the generated code to the Falcon and you know immediately whether it is right — values plausible, units correct. Wrong? Fix it. The iteration loop is minutes, not days.

what goes in

context/ ├── scan-result.json serial params + station IDs ├── capture-10min.csv decoded frame log ├── tagmap-draft.yaml initial tag mapping ├── hmi-screen-01.png main screen ├── hmi-screen-02.png settings page ├── hmi-notes.md operator notes └── target.md goal: remote monitoring + output report

what comes out (within the hour)

out/ ├── tagmap.yaml confirmed tag map ├── client/ │ ├── mqtt_client.py subscribes to Falcon topics │ └── modbus_tcp.py direct read/write (optional) └── README.md deployment + tag notes # Dashboard / SCADA: AI builds this once the customer's direction is understood, not part of this hour

being straight about it

One hour gets you a prototype, not an acceptance build.

We are not claiming AI replaces an entire SCADA project in an hour. What it replaces is the stretch from nothing to the first thing worth looking at — the most labour-intensive and hardest-to-quote phase.

Usually achievable in an hour

A remote view with correct live values, KPIs and trend charts, an alarm list, and a demo you can put in front of a customer or a manager.

Not achievable in an hour

Security review, write-permission and interlock design, long-term database architecture, formal acceptance documentation, and any control logic that needs commissioning on site.

Any remote write capability should go through risk assessment and allow-list restriction. AI-generated code must be reviewed by an engineer before it goes live.

Field data in.Remote HMI out.AI takes an hour.