WAZIR QORNI ABUD, Backend Developer & Cloud EngineerWAZIRQORNIABUD

Problem

Recognition quality had stagnated in production. The question was whether the model or the operating thresholds were to blame.

Architecture

• Go (Fiber, GORM, Casbin RBAC) with one package per feature
• SCRFD detection, ArcFace embeddings through ONNX Runtime on CPU, CUDA or TensorRT
• PostgreSQL + pgvector, Redis, MinIO encrypted storage
• Traefik v3, Prometheus, Grafana, Loki, Tempo, OpenTelemetry
• API keys, webhooks, audit log, plans and quotas, maintenance mode
• k6 load scripts and runbooks for deploy, backup, incident response

Decisions

• Built an evaluation harness that bypasses the vector index, so model and threshold quality are measured on their own.
• Detection threshold 0.45 → 0.20 after sweeping it against a no-face negative set; matching threshold 0.60 → 0.50.
• Fixed ArcFace input normalization, re-ranked in fp32 after HNSW search on halfvec, upgraded backbone r50 → r100.

Verified outcome

Measured on LFW (2,200 pairs): AUC 0.9928, EER 1.46%. Accuracy at the old threshold rose from 87.9% to 93.0% with r100. Discarded faces dropped from 6,142 to 885 on a celebrity set. Cost: 35.4% false detections on an adversarial car-image set. Latency baselines are templated but not yet filled in, so none are claimed.

Problem

Expenses lived in photos and chats. They needed to land in a real ledger, per business unit, with an audit trail.

Architecture

• Laravel 12 API: chart of accounts, journal transactions, cash entries, transfers, fixed assets
• Business-unit scoping through global scopes and team-aware permissions (Spatie)
• Deletion-request approval flow, activity log, weekly finance reports
• Bot: Hono + Baileys, one WhatsApp socket per bot row, admin panel to configure each
• Receipt photo → AI vision extraction → multipart POST to the ledger API
• Credentials and API keys encrypted at rest with AES-256-GCM
• Companion Android WebView app with FCM push (Kotlin) and a TypeScript frontend

Decisions

• The bot is independent of the ledger: its own repo, database and admin panel, talking to the API only.
• AI endpoint, model and per-bot accounts are configured dynamically, not hard-coded.

Verified outcome

In active development: both repositories were updated in October 2026. No usage figures are published, so none are claimed.

Problem

People will not open a dashboard to log a coffee. They will send a message.

Architecture

• Go + Echo with delivery / usecase / repository / domain layers
• PostgreSQL 16 and Redis 7; versioned golang-migrate migrations
• WhatsApp OTP sign-in, no passwords, over the official Meta Cloud API
• Gemini Flash parses free text and receipts into structured transactions
• Organizations, subscriptions, wallets, debts and receivables
• SvelteKit dashboard; Docker Compose and Coolify deployment notes

Decisions

• Official WhatsApp Cloud API instead of an unofficial socket for this product.
• Passwordless login through WhatsApp OTP, since the user is already in WhatsApp.

Verified outcome

Full stack scaffolded with eight migration pairs and deployment docs. No production numbers are published.

Problem

Customers ask “is my laundry done?”. Staff need a fast cashier and a clear record of who did what.

Architecture

• Laravel + Blade web app
• QR code on each receipt opens a public tracking page, like a parcel tracker
• Tracking shows progress per step and the staff member who handled it
• Cashier flow: customer lookup, member discounts, configurable price rounding

Decisions

• Status tracking is public by QR, so customers never need an account.

Verified outcome

Feature set implemented per the project README. No usage figures are published.

Problem

Writing practice needs fast feedback and a record of progress, while staff need control over roles, prompts and AI behavior without a deploy.

Architecture

• Next.js 16 App Router, React 19, TypeScript; server components render thin pages, client components hold UI state
• Server Actions for every CRUD path, each guarded by requireAuth or requirePermission
• Streaming API routes for grammar checking and chat with the model
• Prisma 7 on Neon serverless PostgreSQL, 19 models and versioned migrations
• NextAuth v5 with database-driven roles, permissions and menus
• Admin screens for AI settings, preprompts, users, roles, chat logs and an activity log

Decisions

• Permissions are enforced in server actions, never only in the client.
• Prompts and AI settings live in the database, so staff change them without redeploying.
• BigInt primary keys are converted to strings at the server boundary.

Verified outcome

Deployed on Vercel with Neon. A “hardening” migration and an eslint/TypeScript setup are in the repo. Usage figures are not published, so none are claimed. Login required.