ArogyaLens — Dental AI Platform

The Challenge

Dental clinics in tier-2 and tier-3 cities lack access to specialist radiologists for routine intraoral image review. ArogyaLens needed to bridge that gap by allowing a clinic's staff to capture intraoral images on a mobile device, route them through AI analysis, and produce a structured diagnostic report — all without requiring on-site specialist involvement. The backend had to handle video frame extraction (FFmpeg), multi-image batch analysis via AWS Rekognition, and generate professionally formatted PDF reports with embedded findings. Beyond the AI pipeline, the platform grew into a comprehensive hospital management system — OPD/IPD bed management, pharmacy, labs, HR/payroll, machine management, and e-commerce — all with full RTL-compatible internationalisation across 13 languages including Hindi, Tamil, Arabic, and Japanese.

Architecture & System Design

Mobile app captures intraoral images and uploads to cloud storage. Backend service processes images through AI analysis pipeline, generates written findings reports, and creates branded PDFs. Web portal displays patient results and appointment history. Authentication and payment processing integrated. Full support for 13 languages with right-to-left script rendering.

The Flutter mobile app handles camera capture with real-time preview, Bluetooth accessory integration, and on-device TFLite pre-screening before upload. The Node.js/Express backend manages the full analysis pipeline: S3 upload → Rekognition label detection → OpenAI-assisted finding narration → PDF generation (html-pdf/pdfkit). Firebase handles authentication and push notifications. The Next.js portal (SSR for SEO) provides patients and clinicians with a calendar-based appointment view, diagnostic history, and full-report download. Payment is handled via Razorpay for premium report unlocks.

Code Walkthrough

3-step walk-through of the production implementation — file paths and intent shown above each block.

1. 01

   Step 1 of 3

   Rekognition with a dental label allowlist

   arogyalens-api/src/services/rekognition.js

   Rekognition returns hundreds of generic labels per image (furniture, lighting, etc.). We narrow it to a curated dental allowlist before anything hits the database, so only clinically meaningful findings are stored and billed.

   javascriptCopy

   async function analyzeIntraoralImage(s3Key) {
       const params = {
           Image: { S3Object: { Bucket: process.env.S3_BUCKET, Name: s3Key } },
           Features: ['GENERAL_LABELS', 'IMAGE_PROPERTIES'],
           Settings: {
               GeneralLabels: {
                   LabelInclusionFilters: DENTAL_LABEL_ALLOWLIST
               }
           }
       };
   
       const { Labels } = await rekognition.detectLabels(params).promise();
   
       return Labels
           .filter(l => l.Confidence > CONFIDENCE_THRESHOLD)
           .map(l => ({ name: l.Name, confidence: l.Confidence.toFixed(1) }));
   }

   Takeaway

   The allowlist isn't a post-filter — it's passed to Rekognition's `LabelInclusionFilters`, so irrelevant categories are never scored at all. Cheaper, faster, cleaner.

2. 02

   Step 2 of 3

   Turning raw labels into patient-facing narrative

   arogyalens-api/src/services/narrate.js

   Raw labels like 'CARIES_MODERATE 87.3%' aren't useful to a patient. An OpenAI call rewrites them at a Grade-8 reading level with a strict prompt: only describe findings from the input list, never invent a diagnosis.

   javascriptCopy

   const NARRATE_SYSTEM_PROMPT = `
   You are a dental hygienist explaining findings to a patient.
   Rules:
   1. Only describe findings present in the input list.
   2. Grade 8 reading level.
   3. Never use the words "diagnosis" or "treatment plan".
   4. End with: "Please consult your dentist to discuss these findings."
   `;
   
   async function narrateFindings(findings) {
       const input = findings
           .map(f => `- ${f.name} (confidence ${f.confidence}%)`)
           .join('\n');
   
       const { choices } = await openai.chat.completions.create({
           model: 'gpt-4o-mini',
           temperature: 0.2,
           messages: [
               { role: 'system', content: NARRATE_SYSTEM_PROMPT },
               { role: 'user', content: `Findings:\n${input}` },
           ],
       });
   
       return choices[0].message.content.trim();
   }

   Takeaway

   Low temperature + a tight system prompt keeps the narration deterministic enough for a medical-adjacent context, without hard-coding boilerplate strings.

3. 03

   Step 3 of 3

   Branded PDF report with embedded findings

   arogyalens-api/src/services/report.js

   The final artefact is a PDF the clinic hands to the patient. We render it server-side from an HTML template so the finding table, narration, and clinic logo all stay consistent with the web report — no drifting designs.

   javascriptCopy

   async function generateReportPdf({ patient, findings, narration, clinic }) {
       const html = await renderTemplate('report.hbs', {
           patient,
           findings,
           narration,
           clinic,
           generatedAt: new Date().toISOString(),
       });
   
       const pdf = await htmlPdf.create(html, {
           format: 'A4',
           border: { top: '20mm', right: '15mm', bottom: '25mm', left: '15mm' },
           header: { height: '18mm', contents: clinic.headerHtml },
           footer: {
               height: '15mm',
               contents: { default: clinic.footerHtml },
           },
       });
   
       const key = `reports/${patient.id}/${Date.now()}.pdf`;
       await s3.putObject({
           Bucket: process.env.S3_BUCKET,
           Key: key,
           Body: pdf,
           ContentType: 'application/pdf',
       }).promise();
   
       return key;
   }

   Takeaway

   One Handlebars template, one PDF, one S3 key — the report is a first-class deliverable, not a screenshot of a web page.

Results

ArogyaLens is actively used in pilot dental clinics. The platform supports full-cycle patient journeys from mobile capture to downloaded PDF report. Razorpay integration enables monetisation of premium diagnostic tiers. The Next.js portal's SSR architecture achieves strong SEO scores, driving organic clinic sign-ups. The platform now supports 15+ hospital management modules beyond the original diagnostics scope, with production multi-language support across 13 languages. The admin portal provides cross-department analytics, and the hospital portal handles full OPD/IPD patient lifecycle management.

Gallery & Demos

Admin Dashboard

Overview of all clinic operations: patient appointments, test results, and team performance metrics.

Hospital Portal

Staff view for managing patient beds, treatments, and medical records across the entire hospital.

Multi-Language Support

Language selector showing Tamil and other supported languages for accessibility across regions.

Patient Bed Management Overview

Real-time view of occupied and available beds in OPD (outpatient) and IPD (inpatient) departments.

Multi-Language IPD View

Inpatient department management interface showing patient details in Hindi and other local languages.

Patient Consultation

Doctor view showing patient history, test results, and treatment plan during consultation.

Medicine Management

Pharmacy interface to add and manage medicines prescribed to patients.

Click any image or video to expand · ← → keys navigate

Live Demos