One day inside a practice where every model in the master plan is live and routine.
Told from three seats in the same operatory: the dentist, the hygienist, and the owner.
The committed specification deliberately ships a narrow slice — 2D bitewings and periapicals, a fraction of the source stories, no CBCT, no orthodontics, no generative narrator. That is the correct, fundable first step. This page shows the other side of that roadmap: the complete diagnostic-AI product the architecture is built to grow into.
Chanieldx is the imaging-AI layer only. In practice it plugs into whatever practice-management system a clinic already runs rather than replacing it. That system's own registration, vitals, triage, charting, pharmacy and billing are a separate concern.
It is written as a narrative, not a specification. Nothing here is a commitment — treat it as a vision exercise for where the architecture could lead.
Same operatory. Same practice. Three people, three versions of the same day.
The day's worklist is already triaged: three patients flagged by the predictive risk model as likely to need same-day treatment, ranked above routine recalls. Elena skims it over coffee instead of building it from memory.
The core loop from the MVP is still exactly here, just older and quieter: F1 numbers every tooth, F2 flags a lesion and stages it toward dentin, F3 reports bone level and AAP stage. She barely notices it anymore — which was always the point.
A conversational layer sits over the structured findings now — not deciding anything, just narrating: “Tooth 30, distal surface, roughly 60% through to the DEJ, consistent with the last two visits' trend. No periapical involvement on today's PA.” Elena still makes the call; she just doesn't have to assemble the picture herself.
For a new ortho consult, a low-dose CBCT and cephalometric analysis run through a sibling model to the 2D detectors — same shared-backbone philosophy, new modality. It suggests an appliance window; Elena treats it as one more opinion in the room, same as the 2D flags always were.
A periapical radiolucency shows up on a routine PA for a patient with zero symptoms — the same detector family from the MVP, just tuned over years of data toward the threshold that catches things early instead of late.
Every one of today's flags is still logged as a suggestion the dentist accepted, edited, or dismissed — concurrent-read, unchanged from the MVP's regulatory posture on day one.
None of the future-leap moments above come from nowhere. Each is a straight-line extension of a design choice already made.
A side-by-side, so the vision never gets confused with the plan.
| In the story | Status today |
|---|---|
| 2D caries, bone-level and periapical detection | The actual v1 MVP centrepiece — F2, F3, F4-2D. Shipping, not speculative. |
| Longitudinal before/after tracking | Also real — F12b, called out in the spec as the differentiator. |
| Image-QA retake, PHI redaction, voice charting | Retake (F11) and redaction (F12a) are real v0 stories. Voice-driven charting is a future leap — an imagined interface layer, not specced. |
| Conversational LLM co-pilot narrating findings | Explicitly named in the master plan as post-MVP and out of scope for this build. Future leap. |
| 3D/CBCT and cephalometric analysis | Explicitly out of scope for the MVP. Future leap. |
| Predictive same-day triage, licensable population insights | Not described anywhere in the spec. Invented for the exercise, to test where the architecture could lead. |
| PCCP-based updates, ongoing drift monitoring | Both are real MVP commitments — only their routine, ongoing use is imagined. |
Bottom line: the committed specification is the only part of this that is a build commitment. Everything beyond it is a thought experiment meant to test whether the architecture — shared backbone, structured findings, PCCP, longitudinal registration — actually points somewhere worth building toward.
Back to what is real
Where anything here conflicts with the specification, the specification wins.