Westbridge Materials Pty Ltd
Tax invoice
INV-4628
12 September 2026
Supply of paving and sand
Total AUD 4,280.00
GST included 389.09
PO reference not on page 1
PrimeHQ · Application engineering
Build the software
your business actually needs.
Weeks, not quarters.
Custom web applications delivered through AI-assisted development and a Cloudflare-native architecture. Built to launch quickly, scale naturally and keep evolving as the business changes.
Custom web applications, built faster with AI, engineered for continuous change.
- Launch measured in weeks
- A person still approves the release
- AI inside the app only where it removes work
Example request path
Workers run the application itself: routes, checks, auth decisions, and calls to the data services below. This replaces a traditional application server for the request path.
One possible shape. An application uses the services it actually needs.
Short-cycle development
Software should evolve at the speed of the business using it.
The point is the next release, not a single handover. Each pass is smaller than a traditional project, and a person still decides what is allowed into production.
- 1
Understand
The users, the real steps, the constraints, and what a finished job looks like.
- 2
Prototype
A small working slice that a person can click, before the rest of the system exists.
- 3
Build
The smallest production application that completes the job, including the unglamorous edges.
- 4
Test
Behaviour, permissions, and failure. Generated tests are read, not trusted on arrival.
- 5
Deploy
A staged release. A person decides when it is allowed to face real users.
- 6
Learn
What people actually do, where the process stalls, and which assumption was wrong.
- 7
Improve
The next change, smaller than the first, shipped the same way.
Applications
Turn complex processes into software.
The work is a business process with users, records and a result. Named case studies are not required to see the kind of application. The longer accounts are custom web applications, AI inside the product, and the Cloudflare architecture.
-
Internal business applications
A single place for a process that currently lives in a spreadsheet and an inbox.
-
Operational dashboards
The work in flight, not a wall of charts detached from the queue that produced them.
-
Customer portals
A customer can see status, documents and requests without emailing to ask.
-
Member portals
Accounts, bookings, invoices and records for a membership or a program.
-
Workflow automation
A known sequence of steps, with a person where the decision is not safe to automate.
-
SaaS platforms
A product used by more than one organisation, with tenant data kept apart.
-
Booking and scheduling
Availability, bookings and the exceptions that a calendar link does not handle.
-
Transaction systems
Records of money, stock or entitlements where the state of each item has to be clear.
-
Document workflows
Intake, extraction, review and filing, with the original file kept beside the fields.
-
Reporting and analytics
Figures the business already defines, drawn from the application that created them.
-
APIs and integrations
A clear boundary to Xero, a CRM, a line-of-business system, or another PrimeHQ product.
-
Applications with AI features
Search, extraction or drafting inside a workflow that still has an owner.
An operations application, not a chart template.
A working internal tool keeps the queue, the exceptions and the last run in one place. This is a conceptual member-operations console.
Example interface. Not a client project.
Operations · 1–28 Sep 2026
Overview
Production
Documents processed
All automation runs this month.
Most of the extra time is on documents with no purchase order on the first page.
- 22 supplier invoices this week
- 6 missing a purchase order
- Median handling 11 min, against 9 min last week
| Run | State | Duration | When |
|---|---|---|---|
| Supplier invoice intake | Completed | 2 min | 28 Sep, 14:06 |
| Member document pack | Awaiting review | — | 28 Sep, 13:41 |
| Booking reminder | Completed | 18 sec | 28 Sep, 13:12 |
| Nightly reconciliation | Retry 1 | 4 min | 28 Sep, 02:10 |
Workflows
18 active. Invoice intake, member packs, booking reminders and the nightly reconciliation are the ones with traffic this week.
Documents
- INV-4628.pdf · Westbridge Materials · awaiting approval
- member-handbook-2026.pdf · published
- export-sep-2026.csv · generated 28 Sep
Approvals
- INV-4628 · Westbridge Materials · $4,280.00 · review the purchase order
- EXP-2219 · Site travel · $186.40 · waiting on operations
Reports
The Monday operations summary is generated from D1. It is not typed into a slide.
The same engineering, on the customer side.
A member should be able to find a booking, a document and an invoice without writing in.
Example interface. Not a client project.
Member since March 2024
Jordan Hale
Active
Documents
- Membership handbook — leave and bookings
- Tax invoice 1842 — September tuition
- Court booking policy
Nothing in this example matches that search.
Bookings
Thu 2 Oct 2026, 9:00 · Court 2 · confirmed.
Sat 11 Oct 2026, 8:00 · Court 1 · waitlist.
Invoices
1842 · $186.00 · due 6 Oct 2026 · unpaid.
1790 · $186.00 · paid 4 Sep 2026.
Requests
SR-1044 · Change of contact email · received 26 Sep · in review.
Messages
27 Sep · “Your Thursday court is confirmed.” No reply needed.
AI-assisted engineering
AI shortens the work. It does not take the decision.
Research, drafts, tests and notes move faster. Architecture, security, data access and the production release stay with an engineer.
-
Requirements
Write down the process, the users and the outcome before generating code.
-
Research
Read the existing systems, the constraints, and the platform documentation that applies.
-
Architecture
Choose the smallest set of Cloudflare services that can hold the application.
-
Implementation
AI drafts code and changes. A person reviews the diff before it is kept.
-
Testing
Tests are generated and then edited. Passing tests do not replace a look at the behaviour.
-
Review
Data access, auth, failure and abuse get a human review on every release.
-
Documentation
Decisions and operator notes are written so the next change has a starting point.
-
Regression
The last known behaviour is compared with the change, especially around permissions and money.
-
Iteration
The next release is scoped from what the running application showed, not from a new slide deck.
A review surface, not an autonomous deploy.
The assistant can explain, draft a test, or propose a check. A person accepts the change. The example never ships it.
Example interface. Not a client project.
member-portal
// Active invoices for one organisation. Archived rows stay out.
export async function listInvoices(db: D1Database, orgId: string) {
return db
.prepare(`SELECT id, supplier, amount_cents, status
FROM invoices
WHERE org_id = ? AND archived_at IS NULL
ORDER BY issued_on DESC
LIMIT 40`)
.bind(orgId)
.all()
}export default {
async fetch(request: Request, env: Env) {
const orgId = await requireOrg(request, env)
const rows = await listInvoices(env.DB, orgId)
return Response.json({ invoices: rows.results })
}
}// Wait for a person. Do not mark the invoice paid here.
await step.waitForEvent('manager-review', {
type: 'approval',
timeout: '5 days',
})
await step.do('record decision', async () => {
return env.DB.prepare('UPDATE invoices SET status = ? WHERE id = ?')
.bind('approved', invoiceId).run()
}){
"name": "member-portal",
"d1_databases": [{ "binding": "DB" }],
"r2_buckets": [{ "binding": "FILES" }],
"queues": { "producers": [{ "binding": "JOBS" }] }
}Build passed · 12 tests passed · not deployed
Engineering assistant
The query returns the latest 40 live invoices for one organisation. Archived rows are excluded in SQL, not in the page.
A useful test covers an empty organisation, an archived row that must not appear, and a second organisation that must not leak.
+ if (status && !['open', 'review', 'paid'].includes(status)) {
+ return Response.json({ error: 'Unknown status' }, { status: 400 })
+ }Inside the application
AI where it actually removes work.
The model sees a bounded set of records and returns something the workflow can use: a field, a source, a draft, or a flag. It does not become the owner of the process.
-
Semantic search
Find a passage by meaning, then show the document it came from.
-
Natural-language questions
Ask a question in English and get a table, not only a paragraph.
-
Document extraction
Pull fields from an invoice or a form into the workflow that already expects them.
-
Document classification
Route a file to the right queue, and leave odd files for a person.
-
Summarisation
Shorten a thread or a document for someone who still needs the source.
-
Workflow routing
Suggest the next team from the content, without skipping an approval the business requires.
-
Recommendations
Surface a likely next action from similar past work.
-
Anomaly notes
Flag a figure that moved, with the rows that explain it.
-
Report drafts
Assemble a report from data the application already stores.
-
Knowledge retrieval
Answer from a named set of documents, and say when that set does not contain the answer.
-
Assistants
A bounded helper for one job, with the tools it is allowed to call written down.
-
Bounded agents
A short sequence of steps inside a limit, stopped when a step needs authority the agent does not have.
-
Approval-gated AI
The model proposes. A person approves before a record changes or a message is sent.
A question with sources, not a chatbot.
The useful version shows which records it read, the table it produced, and what it will not do next.
Example interface. Not a client project.
Read invoices and product categories · September 2026
September sales by category
| Category | Amount | Invoices |
|---|---|---|
| Materials | $48,200 | 16 |
| Services | $31,440 | 9 |
| Membership | $18,600 | 100 |
| Hire | $6,240 | 11 |
Hire is down $2,110 on August. The other categories are within 6% of last month. Source: paid invoices in D1, not a forecast.
Task drafted for operations: review Hire invoices. Nothing is assigned until a person confirms it.
Extraction proposes. A person confirms.
The file stays visible beside the fields. A low-confidence value is a review, not a silent write.
Example interface. Not a client project.
INV-4628.pdf · page 1 of 2
- Uploaded
- Text extracted
- Supplier identified
- Invoice number extracted
- Amount extracted
- Categorised
- Awaiting approval
- Record updated
- Notification sent
SupplierWestbridge Materials Pty LtdHigh
Invoice numberINV-4628High
Invoice date12 Sep 2026High
Amount$4,280.00High
GST$389.09High
Needs review · not printed on page 1
CategoryMaterialsMedium
Waiting on a person before the record changes.
A process that can wait without a server waiting with it.
Each stage names the Cloudflare primitive that would carry it. The person at the review gate is still a person.
Example interface. Not a client project.
INV-4628 · elapsed 6 h 12 m · retries 0 · example captured 28 Sep 2026
Submission
Worker
The supplier invoice arrives on a Worker route and is stored as a row plus the original file.
Validation
Worker
Required fields are checked in code. A missing total fails here, before any model is called.
Document processing
Workers AI · R2
The PDF stays in R2. Extraction reads it and writes proposed fields. Low confidence is flagged, not overwritten later.
Manager review
Workflow wait
The workflow waits. No server is held open. Yuki Tanaka in Operations is the named reviewer.
Responsible: Yuki Tanaka, Operations. The run is waiting, not failed.
Approval
Workflow
Approval is an event on the workflow. The model does not approve.
External update
Queue
The accounting update is queued so a slow partner API does not sit inside the approval request.
Notification
Queue
The supplier contact is emailed after the record write succeeds.
Complete
D1
The invoice row in D1 is the record. The PDF remains in R2 for the audit trail.
Cloudflare application stack
A modern stack for modern applications.
The diagram is a set of primitives, not a logo wall. PrimeLabs picks the ones the application needs. Deeper notes on each primitive are planned as their own pages and are not published yet.
What it is
The interface people use: pages, dashboards, portals and installable web apps.
When PrimeLabs uses it
When the work needs a screen, a form, or a dense operational view.
When PrimeLabs does not
It is not the system of record. Business rules that must hold on every request live in the Worker.
What it is
The application runtime. Requests, validation, permissions and integrations run here.
When PrimeLabs uses it
For the request path and for jobs that are still short-lived code.
When PrimeLabs does not
Not a place for multi-day waits or for files. Those belong to Workflows and R2.
What it is
SQL tables for the records the application owns and queries together.
When PrimeLabs uses it
Accounts, invoices, bookings, memberships, workflow records.
When PrimeLabs does not
Not for large files, and not for a hot key that many users must update at the same instant.
What it is
A single place for state that must be consistent for one room, one tenant, or one live document.
When PrimeLabs uses it
Presence, a live editing session, a counter that cannot be loosely updated.
When PrimeLabs does not
Not the default database. Ordinary lists and reports stay in D1.
What it is
Small values read from many places, where a slightly stale read is acceptable.
When PrimeLabs uses it
Configuration flags, a cached lookup, a short-lived token check.
When PrimeLabs does not
Not a relational store, and not where two writes must succeed or fail together.
What it is
Object storage for uploads, originals, exports and generated files.
When PrimeLabs uses it
Invoices, member documents, CSV exports, images the application produced.
When PrimeLabs does not
Not a database. Fields extracted from a file are stored in D1 and point back at the object.
What it is
A buffer between a request and work that can finish later.
When PrimeLabs uses it
Sending mail, calling a slow partner API, smoothing a burst of uploads.
When PrimeLabs does not
Not the engine for a process that waits on a person for days. That is a Workflow.
What it is
A process that can pause, retry, and wait for a person without a server staying online.
When PrimeLabs uses it
Approvals, document pipelines, and anything that must resume after a failure.
When PrimeLabs does not
Not for a single read on a page load.
What it is
Model inference beside the application, for a feature that has a defined input and output.
When PrimeLabs uses it
Classification, extraction, a short summary, an embedding.
When PrimeLabs does not
Not on the path of every request. If the step can be done with rules, it stays rules.
What it is
A vector index for semantic search, retrieval and recommendations.
When PrimeLabs uses it
“Find documents like this” over a corpus the business owns.
When PrimeLabs does not
Not the source of truth. Membership, money and permissions stay in D1.
What it is
A control point for model calls: logs, routing, caching and a place to change provider behaviour.
When PrimeLabs uses it
When the application calls models often enough that failures and cost need to be seen.
When PrimeLabs does not
Not required for an application that does not call a model.
What it is
The network in front of the Worker: TLS, caching, rate controls and DDoS protection.
When PrimeLabs uses it
Every public application. The path is the same one Cloudflare already provides.
When PrimeLabs does not
It does not replace application auth, validation, or a decision about who may see a record.
The application, seen as the services it actually binds.
This is an operations view of one example application. It is not the Cloudflare dashboard, and PrimeLabs did not build Cloudflare.
Example interface. Not a client project.
Application · Member Portal
Service path
Workers · production
D1 · production
86,400 queries in 30 days. Rows for members, bookings and invoices. No file bodies.
R2 · production
Member documents and invoice PDFs. 1.8 GB. Downloads go through the Worker, not a public bucket URL.
Durable Objects · production
One object per live booking board. Used so two staff cannot confirm the same court.
Queues · production
Depth 3. Reminder mail and the accounting handoff. Not the approval wait.
Workflows · production
1,104 runs in 30 days. 7 waiting on a person. 1 retrying a partner timeout.
Workers AI · production
640 requests in 30 days, all from document extraction. Ordinary page views do not call a model.
Vectorize · production
Member handbook and policy pages. Used by the portal search. Invoices are still found by number in D1.
Workers · staging
2,140 requests this week. p95 41 ms. Error rate 0.4% on a fixture import, already fixed in review.
D1 · staging
Copy of the schema, not a copy of production members. 3,200 queries this week.
R2 · staging
Fixture PDFs only.
Durable Objects · staging
Booking-board fixture. Two conflicting confirms are the test.
Queues · staging
Depth 0. Consumers point at a log, not at members.
Workflows · staging
46 runs. Used to rehearse the approval wait.
Workers AI · staging
28 extraction runs against fixture invoices.
Vectorize · staging
A small index of the same handbook, rebuilt when the draft changes.
Releases
AI search improvements. Handbook queries return the section title with the passage. No schema change.
Document processing workflow. Extraction writes proposed fields and stops on a missing purchase order.
UI refinements on the member booking card. Not promoted.
Business outcomes
Less infrastructure. More of the application.
-
Less infrastructure
No application-server fleet to patch, size and watch overnight.
-
Faster deployment
A release is a Worker deploy and a migration, not a machine build.
-
Shorter release cycles
A change can ship when it is reviewed, instead of waiting for a quarterly window.
-
Easier iteration
The next improvement starts from the running application and its logs.
-
Scale without a server fleet
Traffic is absorbed by the platform. Capacity planning is not a rack diagram.
-
Security controls in the path
TLS, rate limits and edge policy sit in front of the code. Application rules still live in the code.
-
Infrastructure as code
Bindings, routes and environments are declared and reviewed with the application.
-
Lower operational drag
Fewer moving boxes. The remaining work is the application, the data and the releases.
Business areas
Organised by the work, not by a thin industry page.
-
Operations
Work queues, handoffs and the status of a job.
-
Customer service
What a customer asked, what was promised, and what is still open.
-
Sales
Pipeline context that should not be retyped into a spreadsheet.
-
Finance
Invoices, approvals and the trail back to a source document.
-
Compliance
Who did what, and the record that shows it.
-
Membership
People, entitlements, renewals and the documents they are owed.
-
Data
A defined set of records, with a question interface only where it helps.
-
Reporting
Numbers that match the application, produced on a schedule.
-
Field operations
Jobs, evidence and status for people who are not at a desk.
-
Professional services
Matters, time and deliverables in one record.
-
Digital products
A product surface that can change without a replatform.
-
Workflow automation
A process with a named owner at each gate.
Process modernisation
Still running a critical process through spreadsheets, email and shared folders?
PrimeLabs can turn that process into an application one workflow at a time. The first release is the painful path, not a replacement for every adjacent task.
What is there now
- A workbook with several owners
- Approvals buried in an email thread
- The source PDF in a shared folder, named inconsistently
What the application holds
- One record, with a status
- A named person at the review step
- The file and the extracted fields kept together
Engineering principles
Deliberately simple, and reviewed.
- Type-safe
- TypeScript at the boundary, so a wrong field fails before a user sees it.
- API-first
- The screen and any later integration call the same rules.
- Observable
- A request, a job and a model call can be traced after the fact.
- Least privilege
- A user, a token and a binding can only reach what that job needs.
- Tested
- The paths that move data or money have tests a person has read.
- Reproducible
- A release can be built again from the same source.
- Staged deployment
- Production is not the first place a change runs.
- Secure defaults
- Closed unless there is a reason to open it. Secrets stay out of the browser and out of git.
- Deliberately simple
- Fewer services, fewer concepts, a diagram a new engineer can still read.
- AI only where it adds value
- A rule, a query or a form wins when the outcome is already determined.
Engagement
Frame, build, operate, evolve.
-
Frame
Understand the users, the workflow, the constraints and the outcome.
-
Build
Create the smallest coherent production application.
-
Operate
Deploy it safely and watch real usage.
-
Evolve
Keep changing the application as the work changes.
Contact
Have a process that should be an app?
Write to PrimeLabs . Say who uses the process, what the finished job is, and what must not change. There is no form and no account on this site.