Cloud-native Architecture
Context
The situation before the platform
Why a DentalTech workflow needed AI rather than a better camera and a bigger spreadsheet.
| The business |
Dental and aesthetic clinics whose clinicians analyse facial and dental structures for smile design and aesthetic treatment planning, and who have to explain those plans to patients in the room. |
| The starting point |
Traditional smile design depended on manual measurements and visual interpretation. Existing imaging systems lacked the intelligence and reliability needed for real-time aesthetic analysis, so evaluations were inconsistent and treatment planning unpredictable. |
| The trigger |
The client aimed to leverage DentalTech AI to modernise clinical workflows, enhance diagnostic accuracy, streamline operations, and ensure consistent outcomes across multiple clinical environments — not just within a single clinic, but at scale. |
| What they wanted |
One intelligent imaging ecosystem that converts static photos and multi-angle inputs into structured, clinically actionable insight, with realistic smile simulations a patient can understand without a technical explanation. |
| Constraints |
Clinicians needed instant feedback mid-consultation, so model latency was a product constraint · camera quality, lighting, patient positioning and viewing angle vary from clinic to clinic · reflections, occlusions and changing facial expressions make teeth, gingiva, lips and landmarks hard to detect · the system had to run across multiple clinics, devices and imaging systems and stay open to future expansion. |
System
What it runs at today
The platform as delivered, live and scaling across clinics and devices.
Revenue per patient
15%
Increase in average revenue per patient after AI-assisted visualisation entered the consultation
AI capabilities
5
Image analysis, treatment planning, smile simulation, patient pipeline management and a conversational clinical assistant
Engagement length
48+
Months, with a dedicated product team through build and scaling
Team size
4+
AI/ML engineers, frontend and backend developers and QA
The engineering problem
Five gaps that shaped the build from day one
Not vague pain points — the specific tensions in running clinical-grade computer vision inside a live patient consultation, each paired with what we did about it.
0
1
Manual and subjective assessments
Traditional smile design workflows depended heavily on manual measurements and visual interpretation. That produced inconsistent evaluations between clinicians and unpredictable treatment planning, with no repeatable baseline to compare a case against.
What we did
Intelligent facial and dental measurement algorithms that automatically analyse proportions, symmetry, tooth positioning, smile lines and facial harmony, so the assessment is data-driven rather than eyeballed.
0
2
Precision on facial and dental structures
Accurately identifying teeth, gingiva, lips and facial landmarks was difficult because of reflections, occlusions, varying facial expressions and genuinely complex anatomy. Small detection errors propagate straight into the measurements built on top of them.
What we did
Custom CNN-based segmentation models isolating teeth, lips, gingiva, eyes and facial regions at pixel level, combined with automated landmark detection and 3D pose estimation for alignment and orientation correction.
0
3
Inconsistent results across devices and environments
Variations in camera quality, lighting conditions, patient positioning and viewing angles degraded feature detection accuracy and overall reliability. A model that only works in a well-lit studio is not a clinical tool.
What we did
A training pipeline on diverse dental and facial imaging datasets, hardened with aggressive augmentation — synthetic lighting variation, motion blur simulation and artificial occlusion — then validated across multiple devices, lighting conditions and facial variations.
0
4
Real-time processing limitations
Clinicians needed instant visual feedback while the patient was still in the chair, but complex deep learning models introduced latency and performance bottlenecks. Analysis that arrives after the consultation ends changes nothing.
What we did
Inference pipelines optimised with CUDA GPU acceleration, ONNX runtime, quantization and model pruning, so analysis returns during the consultation without giving up precision.
0
5
Scaling across clinics and imaging systems
The platform had to work across multiple clinics, devices and imaging systems while staying flexible enough to absorb future AI capabilities rather than being rebuilt for each of them.
What we did
A modular, device-agnostic architecture built for scalable deployment, with integration points for imaging tools, third-party systems and further AI capabilities.
Architecture
How it fits together
Simplified — the shape rather than every service.
0
1
Clients
React.js · Next.js · Responsive web
A clinician-first interface built in React.js and Next.js, responsive across desktops, tablets and mobile so the same visualization works in the surgery and in a remote consultation.
0
2
Gateway
PHP · application layer · Secure data processing · Integration endpoints
A PHP application layer fronts the platform, handling secure data processing and the integration endpoints that imaging tools and third-party systems connect through.
0
3
Services
PyTorch · TensorFlow · ONNX runtime · Pose estimation algorithms · Smile simulation
Segmentation, landmark detection, 3D pose estimation and smile simulation each run as their own inference concern, trained in PyTorch and TensorFlow and served through ONNX runtime so one model can be retrained without disturbing the rest.
0
4
Data & infra
MySQL · AWS · AWS EC2 · CUDA · GPU acceleration
MySQL holds patient imaging history and clinical records, with the platform deployed on AWS EC2 and GPU acceleration behind the inference path as clinic and image volume grow.
Real-time performance was treated as an architectural requirement, not a tuning exercise: GPU acceleration, quantization, pruning and ONNX runtime are what let a complex segmentation and pose-estimation pipeline return an answer while the patient is still in the chair.
Solutions
Six systems doing the actual work
Not a features list — the specific things we built behind every number above.
High-precision semantic segmentation
Custom CNN-based deep learning models isolate teeth, lips, gingiva, eyes and facial regions with pixel-level precision, giving structural analysis something exact to work from.
Landmark detection & 3D pose estimation
Automated landmark detection and pose estimation handle facial alignment, symmetry evaluation and orientation correction, so a simulation holds up even when the photo is not perfectly square.
Real-time AI inference
Inference pipelines optimised with CUDA GPU acceleration, ONNX runtime, quantization and model pruning deliver analysis instantly during a consultation.
AI-driven smile simulation
Intelligent smile previews and treatment simulations show a patient what an aesthetic outcome could look like, which is what turns a plan into an accepted case.
Modular, device-agnostic architecture
A scalable and flexible architecture supporting deployment across clinics, imaging devices and healthcare systems without a bespoke build each time.
AI clinical planning
Measurement algorithms analyse proportions, symmetry, tooth positioning, smile lines and facial harmony, cutting manual assessment time and making decisions more consistent between clinicians.
Key features
What the platform does day to day
Five AI capabilities running in the background of a normal clinic day.
| Capability | Runs | Refresh | What it does |
|---|---|---|---|
| Automated image analysis & insights | On upload | Instant | Processes dental and facial images into segmentation maps and structured clinical summaries |
| Smart treatment planning | Per case | Real time | Landmark detection and 3D analysis automate symmetry, proportion and smile evaluation |
| AI smile simulation | In consultation | Instant | Generates smile design previews that improve patient understanding and case acceptance |
| Patient pipeline management | Continuous | Per stage | Tracks imaging history, monitors progress and flags anomalies across treatment stages |
| Conversational clinical assistant | On query | On demand | Natural-language access to records, insights and diagnostics through one chat interface |
Integrations
How the moving parts plug in
Imaging inputs, inference and clinical records meet in one modular platform layer rather than sitting beside each other as separate tools.
Connected capabilities
Imaging devices & systems
Multi-angle capture across clinics
GPU inference runtime
CUDA acceleration, ONNX runtime
Cloud infrastructure
AWS EC2 hosting and delivery
Third-party clinical tools
Integration-ready module contracts
↓
THROUGH ONE MODULAR PLATFORM LAYER ↓
Platform integration layer
Computer vision pipeline
One processing path for every image
Real-time optimisation engine
Quantization, pruning, fast response
Secure data processing
Encrypted handling of patient imagery
↓
INTO THE CORE SERVICES ↓
Core services
Segmentation
Landmarks & pose
Smile simulation
Patient pipeline
Clinical assistant
Because the architecture is modular and integration-ready, a new imaging device, a new clinic or a further AI capability attaches to the existing pipeline instead of forcing a separate deployment.
Security
What protects patient imagery and records
The platform handles facial photography, imaging history and clinical records, so secure processing and reliable infrastructure were part of the design brief.
Secure data processing & encryption
Patient images and derived clinical data move through secure processing and encryption workflows delivered as part of the engagement scope.
Secure, scalable infrastructure
The infrastructure phase was designed for scalability and performance together, allowing expansion across multiple clinics and users without loosening controls.
Reliability under real-world variation
Augmentation with synthetic lighting variation, motion blur and artificial occlusion, plus testing across devices and facial variations, keeps output stable rather than only accurate in ideal conditions.
Controlled integration surface
Imaging tools and third-party systems connect through defined module contracts, so new connections extend the platform without opening ad hoc routes to clinical data.
Process
How we got there
Five stages, starting with clinical workflow research rather than a model architecture.
0
1
Research & discovery
We worked closely with the client to understand existing workflows, clinical pain points, imaging requirements and the level of precision aesthetic analysis actually demands.
0
2
Data collection & model training
A robust AI training pipeline was built on diverse dental and facial imaging datasets to hold accuracy up in real-world scenarios rather than curated ones.
0
3
AI model optimisation
Deep learning models were optimised with GPU acceleration, quantization, pruning and ONNX runtime to reach real-time inference without giving up precision.
0
4
Frontend & visualisation
A responsive, clinician-first interface was designed so smile simulations, facial proportions and treatment outcomes read clearly in real time. Every UI decision was validated with practising dentists.
0
5
Testing & clinical validation
Extensive testing across multiple devices, lighting conditions and facial variations to confirm stable performance and reliable analysis before release.
Business impact
What changed for the business
Beyond the headline numbers, three things clinicians and patients noticed first.
Consultations that move faster
Real-time AI analysis cut manual assessment effort and streamlined diagnostic workflows, so preparation stopped eating into consultation time.
Less subjective interpretation
Advanced computer vision models improved clinical accuracy by replacing visual judgement with consistent, repeatable treatment insight.
Patients who can see the plan
Interactive smile simulations let patients visualise potential outcomes for themselves, which raised engagement and confidence during the consultation.
Commercial outcome
What the engineering choices are worth in operating terms
Every headline number traces back to a specific decision, not a vague platform effect.
| Engineering decision | Operating outcome | Measured effect |
|---|---|---|
| Pixel-level CNN segmentation | Structural analysis rests on exact boundaries for teeth, gingiva, lips and facial regions | 98% segmentation accuracy |
| Automated measurement algorithms | Proportions, symmetry and smile lines are calculated instead of measured by hand | 47% less assessment time |
| GPU-accelerated real-time inference | Clinicians get analysis during the consultation, not after it | 36% more efficient consultations |
| AI smile simulation and visualisation | Patients understand the proposed outcome and commit to the plan | 91% better engagement & acceptance |
Stack
What it's built on
The actual technologies, not feature names with icons attached.
Frontend
- React.js
- Next.js
Backend
- PHP
- MySQL
AI & computer vision
- PyTorch
- TensorFlow
- ONNX AI
- Pose estimation algorithms
Cloud & infra
- AWS
- AWS EC2
Full case study
Get the complete write-up as a PDF
The same content on this page, plus the extended module breakdown and delivery phases, in a single document you can share internally.
- Full imaging and inference architecture
- Research-to-validation process, phase by phase
- Model optimisation and real-time performance decisions
Download the case study
No spam • unsubscribe anytime • we’re here when you need us
More work
Other platform builds
Hospitality
Every stay, extraordinary.
Hospitality group transformed guest experiences through IoT-enabled room automation, reducing service requests by 58%
Hospitality Group transformed Guest Experiences through IoT-enabled Room Automation, reducing Service Requests by 58%
IoT - Hospitality
Read the case study →