I work on the backend, and on what happens after it ships.

A public SVG-based product simulator paired with a role-based CMS that carries an order from quotation through pre-order, production, and delivery — across Adidas staff, retailers, and suppliers, each seeing only their own slice of the workflow.

• Built order creation, editing, duplication, appended orders, status tracking, and delivery workflows across three role-based portals
• Multi-order PDF export and automated technical-document generation with Headless Chromium
• Extended the SVG simulator: part sorting and validation, colour-palette interaction, NG-colour rules, product-code and branch lookup, customized-image download
• Product search, order filtering, VAS and font master screens, and a scheduled daily customer-master import

A regulated multi-tenant platform covering contracts, licences, treatment processes, electronic manifests, and compliance reporting for providers, clients, and operators. My longest-running project, and the one I have contributed the most code to.

• Query performance on a 20+ GB, 50+ table production database — a manifest list from 11.7s to under 1s, licence matching from 2.7s to 1.2s, a report page from ~450ms to ~30ms
• JWNET integration: EDI request files, result processing, progress-date change tracking, status sync, and error handling
• Authentication, email-based two-factor authentication, and hierarchical multi-tenant authorization
• Scheduled manifest generation, deadline notifications, and high-volume CSV/PDF export

A membership platform for a Japanese eyewear association covering members, stores, payments, training, and Prime accreditation. I owned the move off hand-managed EC2 hosting onto a containerized, Terraform-provisioned architecture.

• Containerized the application and provisioned ECS, ECR, ALB, Auto Scaling, IAM, SSM, and CloudWatch entirely with Terraform
• Built the production CI/CD pipeline — image builds, ECR pushes, and ECS deployments via GitHub Actions
• Designed the ECS on EC2 launch template and autoscaling setup, plus scheduled tasks for member email notifications
• CloudWatch Logs Insights dashboards and saved queries for traffic by functional route

Upload a PDF and read it side by side with its translation, laid out on the page rather than dumped as plain text. Scanned documents go through OCR first; an AI assistant answers questions grounded in the document you are reading.

• Layout-preserving translation pipeline — text extraction, paragraph grouping, and per-page baking so the translated page mirrors the original
• OCR for scanned pages, streamed AI translation with live progress and cancellation, and background image translation
• Document chat over SSE with attachments, chat history, Mermaid diagram rendering, and per-model selection
• Permissioned share links, page-range export, and a security pass that closed a document-enumeration hole (UUIDv7 ids) and an unauthenticated page-image route

Speak in one language while listeners hear another, with live transcription, text-to-speech, AI assistance, session history, and subscription billing.

• Architecture across desktop (Tauri/Rust), web (Next.js), and backend (NestJS)
• Real-time speech pipeline with low-latency audio streaming over WebSockets
• Soniox transcription and text-to-speech with secure API-key storage and audio routing
• JWT authentication, session history, and Stripe subscription billing

A mobile companion, backend, and desktop CLI for driving coding sessions, terminals, and source files remotely — including live preview of a dev server still running on the laptop at home.

• HTTP tunnelling and port forwarding over WebSockets, so a local dev server can be previewed securely from the phone
• Real-time sync between node-pty terminals and the mobile UI, with a source editor and rollback
• Secure device pairing, JWT auth, Redis message routing, Android overlays, and push notifications
• Published as an npm CLI, shipped through to v1.0.6 with unit tests around session and file handling

An N5/N4 grammar app following Minna no Nihongo, scheduled with FSRS-6 spaced repetition and a “golden hour” reminder that nudges you right before a point is about to fade.

• FSRS-6 scheduler with a five-level memory scale derived from card stability, plus leech detection and daily targets
• Four exercise types generated from the content set — cloze, multiple choice, sentence scramble, and typing with romaji input
• Timezone-aware local notifications timed to each learner's forgetting curve
• Idempotent content pipeline for 100+ curated N5/N4 lessons, with furigana parsing and rendering