Product
Winze works like a superintendent, not a dashboard: continuously, in small budgeted steps. It is not a dashboard builder, and it is not an alert-rule tool.
Real screens
Every screenshot here is a real screen from a running demo tenant: simulated plant data behind the real edge, ingestion path and model gateway. Not a mockup.
By area
The same evidence, asked in each team’s own words. Pick the part of the operation you run.
01 — Connect your plant
Point it at the OPC-UA server already on site. Download the edge bundle, run one command on a machine inside the plant, and register the server in the portal; its address space is read and the signals are catalogued and classified. Every connector is read-only, rate-limited and logged in full. Nothing inbound opens on your firewall; the edge only calls out, and each download of the edge bundle makes a single-use enrolment code.
Seven ways in
The plant floor
OPC UA live today; MQTT/Sparkplug, Modbus, historians and substation relays built and shown on the simulated mine
How the edge behaves
Platform feeds
Scheduled REST or read-only SQL pulls, or your system pushes rows
Files that arrive
New files on SFTP or an S3-compatible bucket, mapped and checked
A source not yet built
Becomes a reviewed connector spec, not a software release
A spreadsheet or CSV
A model proposes what the columns mean; a person approves
Structured entries
Phone forms, planned stops, shift rosters, stop codes and quotas
What people see and say
A photo, a field observation or a statement, kept with who said it
One rule for all seven
01
Propose
a model suggests
02
Preview
rows shown first
03
Approve
a named person
04
Quarantine
held, not dropped
Shown working on a simulated five-site mine. No connector has yet run against a customer system.
02 — Make sense of messy data
Say what you have in your own words: a lab spreadsheet, a daily export, a system you already run. Winze proposes a mapping, names its reason and the one uncertainty, and saves nothing until a person approves it. A figure resting on a source nobody has vouched for yet stays provisional until someone sets that source as trusted.
03 — Know your plant
Learned from how equipment starts, stops and loads, or read from the hierarchy you already keep in PI AF or Ignition instead of inferred. One tree, one focus control, from the whole company down to one asset: it only ever narrows what the screens show, never a way to see more than a role allows.
04 — Dashboards that explain themselves
Nothing is kept unless every widget can be drawn from real data. Where a figure can’t be drawn yet, the page says what is missing: a run-state signal, a planned maintenance window, a work order, a reason code.
05 — Ask and investigate
Ask about your site in your own words. You see each check as Winze makes it: what it looked at, what came back, and what it tried next when a first attempt found nothing.
Same answers in any MCP client
Claude Desktop · Microsoft Copilot in M365 and Teams · Cursor · GitHub Copilot
Each with its own key that expires, rate-limits and can be revoked any time. The tools ask, Winze answers, and nothing ever writes back to the plant.
Tools, not guesses
About 350
tools the AI can call, in 12 groups
About 350 tools across 23 servers, counted October 2026. The same tools serve chat, Scheduler, investigations and your own AI client. Shown on a simulated mine.
06 — Compliance
Environmental monitors watched against the limits you configure: a 24-hour average over its threshold, the share of readings that were actually valid. A named person decides what it means and records how the regulator was told. Software never notifies a regulator on its own.
Environment & ComplianceData flows one way
Tenant isolation
Enforced at the database; data never moves between customers.
Your region
Data stays in the region you require: Australia today, more as they are built.
Your model provider
You choose who runs the analysis.
07 — Prove it
The model never types a figure. Every number is a tool result you can open and re-run.
All lost hours
Illustrative splitKnown causes, split out first: maintenance’s issues aren’t sold as discoveries
Unexplained
The real finding: what reaches the GM
Under the hood
Every investigation is recorded (hypothesis, method, data, result) and a memory recalls earlier runs and incidents, so ground is never covered twice.
Orchestrator
Reads the question, plans the work, assigns it
Verification
Checks every conclusion independently before you see it
The method is kept
Every investigation is recorded: hypothesis, method, data, result
Earlier runs feed the next
Runs and incidents are recalled, not redone
Reusable templates (planned)
Promoting a good run into a template
Swap the model
Tools
Query, analyse, report, notify, configure. New MCPs plugged in at any time.
Grows with every integration
Memory and history
Investigations, outcomes and decisions, retained and reused across sites.
Compounds with every job
Prompts and skills
Your engineers’ diagnostic playbooks, encoded, versioned and improved over time.
Tuned without touching the model
Same stack, better model
Run the same tools, memory and skills on a newer model and the answers improve with it. No rebuild required.
The model
LLM, swappable
Scheduler
Set a task, such as “check secondary circuit losses”, and it runs on schedule in the same small budgeted steps, as often as every shift. Every run shows its reasoning, the charts it drew, the tools it used, and who it told.
Self-service
Winze staff needed
None
Step 1
Self-service, no sales call.
Step 2
On a demo mine, before you connect anything.
Step 3
Try it on your own data.
Step 4
One command on site, then register the source in the portal.
No site visit, no setup project, no Winze engineer.
Built versus roadmap
Live today on OPC-UA; historians, maintenance and fleet come next.
See it on your own plant. Book a pilot and we’ll run it with you on one site: the unexplained loss on your plant, priced, with its evidence.
Connect one site
OPC-UA, read-only, one command on site
Run it for 90 days
Chat, shift reports and Scheduler tasks
Review the findings together
The unexplained loss, priced, with its evidence
