OEE
87%
▲ 6%
Running
24
of 28
Downtime
2h 15m
▼ 18%
Alerts
3
attn
P1
Palletiser 01 — Running
98%
CP
Case Packer 02 — Warning
74%
RC
Robot Cell 04 — Faulted
32%
Machine Pilot is the factory-floor AI that answers operator questions about your machinery — grounded in your PLC code, plant SOPs and OEM manuals. Every claim is cited. Every safety bypass is refused. Operators get help in seconds, not hours.
When something goes wrong on the line at 2am, your best operator is asleep, the manual is in a drawer somewhere, and the PLC code is a black box. The cost of guess-and-check is measured in lost production, not labour.
Every operator who leaves takes years of plant-specific tribal knowledge with them. New starters fumble through faults their colleagues solved in minutes.
By the time the right person is reachable, the line has been down for an hour. Most plants escalate too late because operators don't want to "wake people up over nothing."
Without a fast, trustworthy second opinion, operators try the same five fixes in different orders. Real fault diagnosis needs the PLC logic, the SOP, and the manual — together.
Machine Pilot doesn't guess. Every answer is built from your machine's own truth, in a strict hierarchy. The PLC wins over the SOP. The SOP wins over the manual. Contradictions are surfaced, not hidden.
L5X export ingested into a tag graph. When the question is "why is this fault bit set?" the answer comes from the rung that sets it — not a generic guess.
Your cleaning, changeover, and recovery procedures. Versioned and retrievable. Plant vocabulary preserved.
Structure-aware chunking of the manuals. Vision-captioned electrical drawings. Indexed by section, page, and component.
The features that matter on a plant floor — citations, safety refusals, multi-source reasoning, audit trails — are first-class. Not bolted on.
Inline citations to section, page, PLC tag, and routine. Operators can verify, supervisors can audit, suppliers can trust.
Hard refusals on guard bypass, interlock defeat, or any safety-rated override — built into the base prompt, not toggleable by configuration.
One question can pull from PLC, SOP, manual, and drawings in a single answer — and flag where they disagree.
Each customer's corpus, system prompt, and model behaviour is configured per tenant. One platform, many plants — no data ever crosses.
Every operator turn, model decision, and source citation is logged. Supplier audit, plant management reporting, regulatory traceability — all out of the box.
Operators interact on the screen they already use. Future: WhatsApp and Slack adapters drop in without changing the engine.
Operators get answers. Maintenance gets context. Plant managers get visibility. Everyone gets the audit trail.
One platform, three tiers, transparent metering. All tiers ship with citation grounding and safety refusals — they're not premium features.
Customer / OEM documents are sensitive. We treat them that way from day one — encryption, tenant isolation, hard safety rules, and Australian hosting where you need it.
Per-tenant API keys with tier scoping. No cookie sessions, no shared credentials, no auth surprises.
Every database row carries a tenant_id. Every query filters by it. Customer A never sees Customer B's documents, fault logs, or system prompt.
The safety refusals are in the base system prompt. Tenant configuration can extend, never weaken. AS 4024 / ISO 13849 aware.
Every operator turn, every model decision, every cited source — logged with timestamps. Supplier and regulator-ready.
Hosted on DigitalOcean Sydney (SYD1). No data leaves the country unless you explicitly extend to a multi-region deployment.
TLS 1.3 for all traffic, AES-256 at rest for conversation logs and ingested documents. Secrets managed in the DO App Platform vault.
Generic chatbots hallucinate. Machine Pilot is wired into your specific PLC, SOPs and OEM manuals — every answer it gives is grounded in your machine's own documentation, with citations. The model is also configured to refuse any request that would compromise safety (e.g. bypassing an interlock), which generic chatbots do not enforce.
The platform runs in a "manuals and SOPs" mode until the PLC export is available, and clearly tells operators when a question would benefit from PLC context that isn't yet ingested. When the L5X export arrives from your OEM, ingestion is a single job run — no code change.
For an existing customer adding a second site: drop the new site's documents into their cloud folder, add a tenant row with site-specific configuration, run the ingestion job. Typically under one hour of work. For a brand-new customer: a kick-off call to map their documentation, then 1–2 weeks to first operator demo with their real corpus.
The system prompt is hardened to require source attribution on every factual claim. If the retrieved chunks don't cover the question, the model is instructed to say so and ask a clarifying question — not to invent fault codes, tag names, or procedures. We monitor citation accuracy as a first-class quality metric.
Yes. The API is the product. The browser / tablet wireframe you see in the demo is one interface; Slack, WhatsApp, and embedded-in-your-own-dashboard integrations all hit the same /ask endpoint. The roadmap covers Slack escalations in Phase 2 and WhatsApp field-tech mode in Phase 3.
You do. The conversation logs and ingested documents belong to you. We use the platform's logs to improve quality on your tenant only — never to train models that other tenants benefit from. Enterprise tier includes a contractually scoped data-use clause.
We're booking 10-minute discovery calls with plant managers and OEMs through May–June. Walk us through one of your machines, see Machine Pilot answer a real question from your documentation, decide if it's worth a pilot.