Trainee mode

Trainee mode is for the engineer learning on the job — usually a junior or improver running a procedure on real work for the first few times. The AI assistant on the glasses actively coaches them, can see what they're looking at (on request), and can fetch information from your knowledge base, the Parts Arena catalogue, and any custom skills you've built.

When to use trainee mode

• A new hire or apprentice doing a procedure for the first time on real customer work.
• An experienced engineer hitting a model of equipment they haven't worked on before — let the AI pull the manual and exploded-view diagrams in their eyeline.
• A field engineer who needs a second opinion on something unfamiliar — gas valve, fault code, wiring puzzle — without calling a senior off another job.

What the engineer experiences

1. Put the glasses on, power them on. The Core App authenticates automatically.
2. Choose the task they're working on (synced from your job-management integration or assigned in the Dashboard).
3. Say "Start training session."
4. The session connects (~2 seconds pre-flight).
5. The AI greets them and is ready to talk. The procedure for the task (if there is one) is loaded as context.
6. The engineer talks naturally as they work. They can ask the AI questions, request manual pages, ask for parts diagrams, or just narrate and have the AI confirm they're on track.
7. When done, "Stop training session."

What the AI can do in trainee mode

In trainee mode the AI uses the full gpt-4o-realtime-preview model with function calling. It can:

Talk in real time

Server-side voice activity detection means the engineer just talks — no command word, no push-to-talk. The AI replies in voice, in the open-ear speakers, and the engineer keeps their hands on the work.

Look at what they're looking at

The engineer can say "look at this" / "what does this say" / "can you see the model plate" — the AI calls a capture_camera skill, grabs a frame from the glasses camera, analyses it, and answers. Useful for:

• Reading fault codes off a tiny boiler display.
• Confirming a part number on a plate the engineer can barely see.
• Identifying which valve they're looking at on a complex manifold.

Search your knowledge base

The AI has a search_knowledge skill that does semantic search across everything your tenant has uploaded — manuals, procedures, wiring diagrams, internal specs. When the engineer asks something specific the AI looks it up first rather than guessing.

The skill description includes a manifest of your specific document names, so the AI knows what's actually available before it tries.

Run platform + custom skills

If your tenant subscribes to bundles that include platform skills (Parts Arena, custom integrations), the AI gets those tools too. For example, with the Parts & Spares bundle active the AI can:

• Identify a boiler from its GC number.
• Search the Parts Arena catalogue for a part by name.
• Pull a manual page or exploded diagram and display it in the engineer's eyeline.
• Voice control: "show me the burner exploded view" / "show me page 12" / "next page."

Skills you've built yourself (see Creating a custom skill) appear in the same way — the AI knows when to call them based on each skill's description.

Voice patterns that work well

Engineers learn quickly, but here are the patterns that consistently get the best results:

• "What does fault code F75 mean?" — direct question, the AI checks the knowledge base + the manual page for the model if one's loaded.
• "Show me the exploded view of the burner." — explicit request for a visual, triggers the right Parts Arena skill if it's entitled.
• "What's the test pressure for this model?" — the AI checks the manual.
• "Look at this — am I in the right place?" — triggers the camera capture.
• "Walk me through the next step." — if a procedure is loaded for the current task, the AI advances through it step by step.

The engineer doesn't need to know which skill they're triggering — they just describe what they want and the AI picks the right tool.

Viewing the session afterwards

Trainee sessions are recorded the same way trainer sessions are. The session detail page in Dashboard → Sessions shows:

• The full POV recording (Video Recording card)
• The AI-generated session summary — score (0–100), prose summary, trainer notes if any (Session Summary card)
• The procedures extracted from the session (Procedures card)
• A Linked Task card if the session was run against an integration job

The page is read-only — there's no step-level audit log, no per-skill timing breakdown, no completion checklist. The Session Status filter on the Sessions list (Completed / In Progress / Failed) is the main triage tool for trainee sessions that didn't go cleanly.

See Viewing a session for the full view side.

What you don't need to do

• No setup per session. Bundles allocated to the tenant become available to every paired pair of glasses; no per-session configuration.
• No selecting the model upfront. The AI works from what the engineer asks. If a procedure is loaded for the task, the AI uses it as a guide; if not, it answers from knowledge and skills as needed.
• No "ending the session early" — sessions just stop when you say to. Minutes are billed for actual session time, not pre-allocated blocks.

When the AI doesn't know

If the AI doesn't have the knowledge to answer (no relevant document in the knowledge base, no skill that can help, ambiguous question), it'll say so directly rather than guess: "I don't have a specification for that model — try the manufacturer's spec sheet" or "I can't see clearly — can you look at the model plate from a slightly different angle?"

This is by design. We tune the trainer/trainee agents to defer rather than hallucinate, especially on safety-critical steps. If you ever see the AI give an answer that's confidently wrong, please flag it via the session review — that's the kind of signal we use to improve.