Krest One Dental · Case study Healthcare Ops · Dental · both halves
Thirty-five mailboxes on one side. A fine-tuned caries model on the other. Every decision routed to a clinician for the final call. Krest One Dental’s paperwork and imaging, on one operator, with a human between the model and the money.
DentalAuto is an end-to-end AI operator for dental clinic automation. It bundles two tightly-integrated halves: email automation across 35+ Microsoft Exchange office mailboxes (classifying inbound messages into six categories — insurance payment, portal login, claim status, patient billing, office admin, spam — and routing AR-relevant emails to the coordinator), and dental vision (a fine-tuned YOLO11m caries detector with a MedGemma 4B-IT annotator running on intraoral photographs). Every decision routes through a human for the final call; confidence thresholds at 0.70 (auto-forward) and 0.50 (review queue floor) are configurable per office.
Discovery interviews with Krest One Dental’s AR coordinator and treating dentist shaped two design rules: the email classifier needed two thresholds (not one) — auto-forward high-confidence insurance emails, queue mid-confidence for review, hold low-confidence silently flagged; and the vision model should annotate (write rationale) rather than veto (make verdicts). The email half uses Claude Haiku 4.5 via OpenRouter for classification with structured JSON output; Supabase Realtime for live dashboard updates; Row-Level Security for authenticated staff access. The vision half trains on 1,846 patient-safe occlusal-only intraoral photographs with zero patient-level overlap between splits, fine-tunes MedGemma 1.5 4B-IT with LoRA (rank 16, α 32, bf16 + 4-bit NF4), and deploys via Ollama GGUF Q4_K_M quantisation at ~1.5 GB runtime.
The email half shipped on Vercel with Next.js 16 + React 19 + Tailwind 4. Microsoft Graph webhooks ingest email from 35+ office mailboxes in under two seconds end-to-end. The dashboard shows email list, detail view with AI reasoning, review queue, forwarding-rules editor and inbox management. The vision half is deployable via Ollama on an M4 Mac mini or cloud K8s. A two-stage pipeline runs YOLO11m for caries detection (bounding boxes), then per-detection crops flow through fine-tuned MedGemma to produce verdict + rationale, composited into a clinician-facing panel.
Confidence thresholds must be tuned through conversation, not guesswork — Krest One’s request for two thresholds (not one binary yes/no) was the origin of the 0.70 / 0.50 design. When we tested MedGemma as a veto filter it killed 21.5% of true positives to save 28.5% false positives, dropping F1 from 0.865 to 0.769. Shipping YOLO alone with MedGemma as a reasoning annotator (not a veto) was the right call. Patient-level splits — not image-level — are non-negotiable for medical-imaging credibility.
Krest One Dental’s AR coordinator stopped triaging 200+ daily emails manually — high-confidence emails auto-forward, mid-confidence ones land in a review queue. The dentist reviewing intraoral photos now sees YOLO detection boxes and MedGemma rationales on every image — a second pair of eyes that leaves a note. No decision is fully autonomous; every action is logged, timestamped and defensible for compliance and QA.
“The AR coordinator asked for two thresholds, not one. That was the origin of the 0.70 / 0.50 design — and the reason the model annotates instead of vetoes.”
Confidence thresholds at 0.70 (auto-forward) and 0.50 (review queue floor), configurable per office. Six email categories: insurance payment, portal login, claim status, patient billing, office admin, spam. Vision fine-tunes MedGemma 1.5 4B-IT with LoRA (rank 16, α 32, bf16 + 4-bit NF4) on 1,846 patient-safe occlusal-only intraoral photographs.
A template-first builder for teachers, event planners and hobbyists. PDF export of up to 1,000 unique cards, live-play up to 2,000 online players, and a global base of 550k+ users. One click from landing page to printable card.
Send a paragraph about what you’re building to hello@creative-mantra.com. We reply within 24 hours with next steps and, for qualified scopes, a fee letter within four working days.