Two products, filmed end to end.

What the customer gets

• Their own bot, not ours. The business's name, its vocabulary and one of ten palettes — every screen a patient sees belongs to the practice. Nobody is sent to a shared directory to find you.
• One permanent link. It goes on the card, in the Instagram bio, into a forwarded message. It never changes, so nothing printed goes stale.
• A panel on the owner's phone. Today's bookings, next week's, cancellations and blocked time — the same Telegram they already keep open, with no separate login.
• Reminders that go out on their own. The message lands on the morning of the appointment and names the time and the address, which is the no-show most practices are actually paying for.

Three decisions worth knowing

• Bookings are intervals, not slots. Once a twenty-minute visit and a sixty-minute consultation share a calendar, guarding the start time protects nothing — 10:00 and 10:10 have different starts and overlap by ten minutes. Overlap is refused by the database itself, so a race between two patients tapping at once cannot double-book anybody.
• One practice cannot read another's patients. Isolation is enforced twice: in the application, and again by row-level security inside Postgres. The second layer is the one that still holds when the first has a bug.
• A late renewal does not take the page down. Payment is card-to-card and an operator has to approve it, so a practice that pays at midnight would otherwise go dark while waiting on us. A four-day window keeps the booking page fully live — new appointments included.

What one photograph produces

• Eleven regions, found without a model. Face landmarks map the forehead, brow, nose, cheeks, the skin around each eye and the mouth, the chin and both jawlines. That step is deterministic, runs on a CPU in well under a second, and costs nothing per scan.
• Six axes, scored and banded. Dryness, surface evenness, redness, tone unevenness, lines and pores — each nought to ten, each landing in one of five named bands rather than a bare number.
• An overlay on the photograph itself. Drawn in a single cool hue on purpose: warm heat colours over skin are almost impossible to tell apart from the redness they are drawn on top of.
• A skin-age number that is allowed to be compared. The model is only asked for a decade, because a model asked for "34" will confidently say 34 today and 29 tomorrow on the same face. The position inside that decade is computed from the scores by a fixed rule — so a re-scan three months later moves only if the skin moved.
• Every earlier scan, kept. The report's job is to be compared with the next one, so the history is the product rather than a feature on top of it.

Why it will not name a disease

• The vision model is forbidden from it. It may describe distribution, colour, border quality and symmetry from a closed vocabulary of what a photograph can show — silvery scale, a depigmented patch, redness with no traceable edge. It has no permission to say the word for any condition.
• Named differentials come from a table, not a model. The step from "these descriptors" to "consider eczema" is a lookup a specialist can read and defend, and it answers the same way twice. A model asked to name conditions will name them, confidently, off a bad JPEG.
• Confidence stops at moderate. There is no "high" band in the code at all — not reserved, not unreachable, absent. One uncontrolled photograph does not earn it.
• Both directions are recorded. Every candidate carries the evidence against it beside the evidence for it. A list that only argues for its own entries is a list of suggestions.

We analyse. The doctor decides.