Dhevan Anthareza

Production SaaS combining physical direct mail with multi-channel digital follow-up — Google geotargeting, social retargeting, YouTube ads, QR codes, call tracking. Built as a whitelabel multi-tenant platform serving hundreds of partners managing campaigns at scale: millions of ad impressions and hundreds of thousands of mailed pieces tracked.

• Migrated file processing to S3 to enable horizontal scaling — refactored local pipeline with backward-compatible rollout, then removed legacy.
• Extracted QR code generation into a Go microservice for performance, communicating via internal commands.
• Integrated Flowcode for branded QR generation and analytics.
• Optimized statistics dashboard performance — caching layer keyed by URL + query parameters cut load from 10s+ to under 1s on heavy filters.
• Automated Zoho billing operations (updates, cancellations) and USPS campaign management via direct API integration.
• Led incremental Laravel upgrades v8 → v12 over three years; separated API/web routes and introduced proper CORS in the v12 migration.

Live consumer-facing mobile app letting contributors across the U.S. scan and submit physical mail pieces in exchange for cash. Captured mail feeds DirectMail2.0's WMW database — the largest AI-powered direct mail intelligence database in the U.S. with 300,000+ catalogued pieces. Handles real-money payouts, government-grade identity verification, and on-device computer vision.

• On-device document scanning with ML Kit (Android) and Vision Kit (iOS) — auto-detects mail piece edges, corrects perspective, enhances captures locally for speed and privacy.
• Stripe Identity for KYC verification, layered with libpostal address normalization to match verified ID address against user-input registration address.
• PayPal payout integration for contributor compensation; resolved tricky permissions/credentials issue by tracing PayPal docs.
• Direct-to-S3 uploads via presigned URLs — backend mints URL, mobile uploads directly, slashing backend transit.
• AWS Cognito for shared authentication across mobile and backend.
• Built the 6-step contributor onboarding flow (Learn → Profile → Test Mail → ID Verification → Method → Approved) end-to-end.

AI platform built specifically for direct mail performance modeling and optimization — publicly launched March 2026. Users upload mail piece creatives and receive AI-powered analysis grounded in DirectMail2.0's dataset of 80,000+ campaigns delivered to nearly 2 billion recipients. Combines proprietary direct mail intelligence with multi-model LLM validation.

• Custom Zoho Hosted Payment Page integration with branded logo, custom webhook handling, and conditional credit-card-on-file logic.
• Zoho customer data sync — designed reconciliation logic mapping platform users to Zoho customer IDs, preventing duplicates.
• Multi-LLM provider integration layer — owns the code routing requests to OpenAI, Gemini, and Claude. Handles failures, malformed responses, cross-provider behavior diffs so the prompt-engineering team works against a stable interface.
• Fixed real-time WebSocket gateway for file processing progress (NestJS Gateway) — squashed duplicate listener registration bug; added 'last message' recovery so reconnecting clients resume seamlessly.
• Multi-step progress UI (steps 1–4 + percentage) during seconds-to-minute file processing.

Production attendance app for staff at Universitas Dian Nuswantoro (Udinus), letting employees clock in from their phones using face verification and on-campus location enforcement. Live on both the App Store and Play Store, with hundreds of university employees as active users. Replaces traditional fingerprint/card-based attendance with a phone-native flow: open the app, verify your face, confirm you're on campus, done.

• Built the entire mobile app in Flutter as the sole mobile developer, integrating with a Laravel backend built by a separate team. Single codebase shipping to both iOS and Android.
• On-device face verification pipeline: Google ML Kit detection (real-time bounding box) → face crop → FaceNet embedding via TensorFlow Lite → Euclidean-distance similarity match against the registered embedding from the backend.
• Privacy by design — only the resulting embedding vector (a plain numeric array) is transmitted to the server, never raw face images. Bandwidth and privacy win in one move.
• Embedding storage strategy — at hundreds-of-users scale, embeddings are stored as float arrays alongside the employee record in the regular database. No specialized vector DB needed; simple, maintainable, good enough.
• GPS-based geofencing — home screen compares device GPS to the registered campus location in real time, surfacing "Di Dalam Area" / "Di Luar Area" status. Attendance is only valid on campus.
• Polished attendance dashboard — profile, live server time, geofence status, and scrollable history with clear Indonesian status badges (Masuk Tepat Waktu / Pulang Tepat Waktu / Izin / Sakit).
• Cross-platform shipping — took the app through both App Store and Play Store review, including camera/location permission justifications and privacy policy requirements specific to face-and-location apps.

Medical AI platform helping clinicians analyze chest scans for tuberculosis detection and anomaly identification. Doctors upload chest imagery (DICOM or standard images); the system returns TB classification, anomaly localization, and a GradCAM heatmap overlay showing why the model made its prediction. Part of the Oncodoc cancer-screening ecosystem with Universitas Dian Nuswantoro.

• DICOM ingestion pipeline — accepts DICOM alongside regular images. Originals archived; converts to JPG for both display and model consumption.
• Image preprocessing for cross-device consistency — different scanner manufacturers produce images with different intensity ranges; built normalization layer ensuring consistent predictions across imaging devices.
• Python inference service with FastAPI — loads ONNX and Keras .h5 models, exposes inference as REST, returns predictions plus base64-encoded GradCAM heatmap.
• Architectural decision: Laravel + Python split. Python isolated for ML; Laravel kept for the main app to keep it maintainable by a broader pool of developers (real bus-factor consideration for a small team).
• Frontend visualization with Nuxt — renders original scan, predictions, and GradCAM overlay returned from the Python service.