Problem
A health-tech client needed an operational platform around health-test kits, not just a public website — and that's what I built with them. Customers had to register kits and follow test status. Stores needed to invite and support customers. Administrators needed to manage exceptions. Lab events had to flow back into the product, and notifications had to be timely without making every request depend on a fragile chain of third-party calls.
Constraints
Customer, store and administrator responsibilities needed distinct access. Lab events could arrive out of order or before a customer record existed. Sensitive communications needed an operator's judgment.
What I did
Domain model
The first engineering task was mapping the real process with the company’s operating team:
- who touches a kit at each stage
- which information is required before the next transition
- which events originate at the lab
- which events notify a customer, store, or administrator
- which cases require review before communication
The core domain object is a test kit identified by a vial number. Around it sit customers, stores, users, media assets, status notes, lab events, notifications, and support exceptions.
The product has three distinct operational surfaces:
Surface | Responsibilities |
|---|---|
Customer | Register kit, verify identity, track status, view results. |
Store | Invite customers, create tests, view history, resend invitations, manage release readiness. |
Admin | Manage stores and users, inspect tests and media, resolve exceptions, review status and notifications. |
Treating store users as a thin version of customer accounts would have produced incorrect ownership boundaries. Store access required organization-style membership, role checks, and store-scoped data access.
Architecture
The Cloudflare-first TypeScript stack kept the operational surface small while providing distinct homes for relational state, media, and asynchronous work.
Layer | Choice | Responsibility |
|---|---|---|
Web application | Next.js on Cloudflare Pages | Public registration, customer, store, and admin interfaces. |
API | Hono Workers with tRPC | Typed procedures, webhooks, uploads, and request validation. |
Relational state | D1 with Drizzle | Customers, stores, tests, status, notifications, and audit data. |
Media | R2 | Vial images and public media assets. |
Background work | Queues | Lab pushes, invitations, transactional email, and retryable side effects. |
Authentication | Clerk | Identity, verification, invitations, and organization-style ownership. |
Resend with React Email | Typed transactional templates and delivery status. |
The monorepo separated the web application, API, worker, database schema, transactional templates, auth, shared types, UI, and integrations. That kept the UI independent from lab processing and kept the worker independent from form rendering.
Asynchronous notification delivery
Email is part of the operational workflow, so delivery needs an audit trail. Product events enqueue jobs rather than sending most messages inside the request path:
product event
-> Cloudflare Queue message
-> worker job
-> React Email template
-> Resend send or draft
-> D1 notification record
-> Resend webhook updates delivery statusThe record captures recipient, provider, message metadata, and delivery state. Retrying a job is safer than retrying an opaque browser request.
Sensitive cases use a reviewed draft. Lost-sample and sample-error flows can alert support and create customer-facing communication for an operator to approve. Automation removes repetition; it does not remove judgment from sensitive messages.
Lab event processing
The lab webhook is intentionally thin. It validates the incoming event and enqueues it. A worker resolves the current vial status, finds the matching test, updates the lifecycle, and triggers the appropriate follow-up.
The design protects terminal states and accounts for out-of-order operations. A late webhook must not move a released test backwards. An event that arrives before the customer record exists should become an inspectable holding case rather than disappearing into logs.
Result
Underneath the dashboards, this was really an operational reliability project. The system ended up with typed application boundaries, durable workflow state, asynchronous integration processing, auditable communication, and enough administrative surface for the business to handle its normal exceptions without pulling in an engineer.
The delivered system connected customer, store and administrator workflows. No clinical outcome or financial impact is claimed here.
Tradeoffs
Asynchronous processing makes integration failures inspectable but requires durable lifecycle state, retry handling and administrative tools. Operator review remains part of sensitive communication workflows.