What you'll do

• Establish measurement. Build per-model, per-customer, per-category accuracy reporting that Product and Sales can act on. Nothing else you do matters until this exists.
• Build the correction flywheel. Capture the quality-control corrections our customers already make, turn them into structured, versioned training data, and close the loop with human review where confidence is low.
• Standardize the squads. One deployment path, one evaluation harness, one model and dataset registry across restoration design, scan quality, prescription extraction and agents.
• Solve multi-tenant model behaviour. Thousands of customer-specific product codes, prescription conventions and unwritten preferences sitting over a shared model base - without building a bespoke model per customer.
• Take inference to production scale. Move workloads from on-premise to cloud with understood cost per inference and real concurrency headroom.

• • Set technical direction for AI across the company, partnering closely with the CTO, Product and the operations teams in Manila who currently absorb the work our models can't yet do.
• • • • Before you apply - how we workThis is a full-time in-office role at our Vancouver office, five days a week. We are not offering remote or hybrid arrangements for this position, and this is not negotiable at offer stage.
We're explicit about it because the work genuinely depends on it. You'll be standardizing how four existing AI squads build, working across a Manila engineering organization, and spending real time with the operations people who currently absorb the work our models can't yet do. That happens in a room.
If you're looking for remote or hybrid, this isn't the right role and we'd rather not waste your time.

• What we're looking for
  Required
• Significant experience building and operating machine learning platforms in production - model registries, versioned datasets, CI for models, monitoring, rollback. Not research infrastructure.
• Depth in document and information extraction from unstructured, inconsistent text. Prescriptions are the core data object here and they are messy.
• You have built a system where user corrections became better models, and you can walk through exactly how the data moved.
• Genuine evaluation rigour - you can define what "accurate" means for a given category, defend it, and ship against it.
• Judgment about when not to use a model. A significant share of this problem is classification and mapping, and deterministic rules beat ML for a lot of it.
• Ability to align engineers across squads and time zones without formal reporting authority.

• Strongly preferred
• Multi-tenant or per-customer model behaviour at scale.
• 3D or geometry-based ML - directly relevant to scan quality and restoration design.
• Experience with healthcare or other regulated data; patient privacy and data residency are live constraints for us.
• Inference cost management and on-premise to cloud migration.

• • Not required
Dental or medical-device background. The workflow is learnable in an afternoon and we would rather have platform depth. What we do want is curiosity about a strange, physical, high-stakes manufacturing process - the output of our software ends up in someone's mouth.

What success looks like

By 90 days - you can tell us, with evidence, where our models actually fail and for which customers. The squads agree on a single evaluation standard.

By six months - customer corrections flow into training data automatically. At least one production model has measurably improved through that loop.

By twelve months - prescription quality control runs with confidence thresholds, routing only genuinely ambiguous cases to human review. We can state our accuracy publicly and defend it in a sales conversation.

Why this role

Larger competitors have bigger AI teams and a five-year head start on modelling. We are not trying to beat them at that.

We are trying to build something they structurally cannot: a system that gets better every time one of six thousand laboratories tells us we got something wrong. That data exists. The loop does not.

If you have built this kind of flywheel before - or you have been close enough to one to know exactly why most of them fail - we should talk.