Halyard

A wind turbine technician spends the working day ninety metres up, inside a nacelle the size of a shipping container, wearing a fall-arrest harness, hearing protection, and gloves rated for electrical work. The machinery is loud. Their hands are on a torque wrench. And yet everything they need — the torque spec for this turbine variant, the isolation checklist, the fault history that would tell them another technician found the same problem four months ago — is on a tablet they cannot safely reach.

The result is a workflow built on interruption: stop, de-glove, retrieve the device, find the information, re-glove, resume. Fourteen times per turbine visit, on average, costing roughly two and a half minutes each — and often skipped entirely, with the technician working from memory instead. My client was building a voice-first agent to remove that interruption. I led research, conversation design, and product design from the acoustic field study through build-ready specification.

• Contextual Inquiry
• Acoustic Field Study
• Conversation Design
• Dialog State Modeling
• Multimodal Wireframes

I spent nine weeks across four sites in Scotland, Germany, and Texas — GWO-certified, harnessed, following the same protocols as the crew — running ride-alongs, nacelle contextual inquiry, and Wizard-of-Oz voice trials where I acted as the agent over radio while technicians worked. A parallel acoustic field study measured ambient noise at nine work positions across three turbine models, because the entire product direction depended on whether speech recognition could function in that environment.

The Wizard-of-Oz trials surfaced the design risk that shaped everything downstream: technicians respond to a read-aloud safety instruction with "yeah," "mm-hm," or a grunt — while still mid-step. A system that accepts that as confirmation will eventually advance a lockout procedure past an incomplete isolation. Every safety pattern in this project traces back to that one finding.

Fourteen participants — technicians, site leads, and O&M managers — across ride-along observation, nacelle contextual inquiry, calibrated acoustic measurement, and eighteen Wizard-of-Oz voice sessions.

Three roles across the operational hierarchy. Ewan is the primary user — Halyard is designed for him first. Marisol governs the knowledge layer it creates. Thomas holds the safety accountability that determines whether it can be adopted in the field.

The brief could have grown into a general-purpose maintenance platform. I scoped it down to the interaction that actually removes the de-gloving tax, and deferred everything adjacent.

Turbine control was cut entirely — Halyard is advisory only, and any action affecting machine state runs through certified controls the technician operates directly. That boundary isn't a UX nicety; it's what makes the product approvable at all for a buyer like Thomas, whose primary objection is exactly this risk.

Predictive analytics, blade inspection and drone imagery, and offshore operations were deferred as different workflows for different users. Scheduling and dispatch stay with the existing CMMS — Halyard receives the work order, it doesn't create or optimize it. And autonomous report submission was never on the table: the technician always reviews and signs, and no report leaves the device unsigned.

• Cut: Turbine Control
• Deferred: Predictive Analytics
• Deferred: Offshore Operations
• Kept: Guided Procedure + Diagnostic Reasoning

Three functions, in priority order — documentation is deliberately third, a byproduct of the first two rather than a task the technician performs. Guided procedure execution reads torque specs, sequences, and safety steps aloud at the technician's pace, confirming critical steps explicitly and never advancing on a timer. Diagnostic reasoning lets the technician describe symptoms and hear likely causes ranked by probability, informed by fleet history — including what a specific previous technician found on this specific machine, with attribution spoken aloud.

Underneath both sits a register model that shifts from conversational to terse the moment the interaction crosses a safety boundary — sentences shorten, structure becomes numbered, and the confirmation requirement changes from any casual acknowledgement to one fixed phrase. The shift is itself a signal to the technician that the stakes have changed, backed by a two-tone earcon at the boundary and a 0.95 confidence floor with no margin for interpretation.

A voice-primary product has no navigation the technician can browse or recover a wrong turn from — the architecture lives entirely in what the system can understand and what it holds in context. Thirty-one intents organize into six families, each with its own confidence floor and failure behaviour calibrated to the cost of being wrong, not to a single global accuracy target.

Four states from the dialog model, traced through a single turbine visit — from the moment Ewan wakes Halyard at the tower base to the safety gate that justifies the whole product.