Inspiration
Project boards say what was planned. Invoices say what is being charged. The difficult decision between them is often left to scattered notes: what evidence shows that a milestone is ready to bill?
Milestone Desk gives freelancers and small studios one place to inspect that evidence, record a decision and prepare the corresponding invoice. The goal is a review that another person can understand, rather than an AI-generated assertion that the work is complete.
What it does
A genuine AG Grid Community table compares milestones, fees, review status and AI evidence coverage. Selecting a milestone opens its acceptance criteria and supplied delivery evidence.
The AI reviewer maps exact source excerpts to each criterion, identifying direct support, partial support and gaps. Server code checks that the excerpts actually exist in the source. The person then records a decision and reason for every criterion. AI suggestions never count as human acceptance, and changing the evidence invalidates the previous review and acceptance.
Only a fully accepted milestone unlocks the PayPal sandbox invoice preview. The person inspects the recipient, amount and exact payload before creating an unsent draft. The server verifies the returned invoice and keeps a durable attempt record so concurrent or repeated clicks do not create duplicate drafts. The activity journal and exportable review packet explain the path from evidence to acceptance to invoice.
Both light and dark appearances cover the comparison grid, evidence workspace and review dialogs. The interface follows the system initially and remembers an explicit preference.
How we built it
Milestone Desk uses Node.js, SQLite, native JavaScript and a pinned, locally bundled AG Grid Community 36.3.0. The grid provides real row selection, sorting and search for the milestone comparison workflow.
A configurable OpenAI-compatible HTTP endpoint performs structured evidence review. The free local setup has been verified with Ollama and qwen2.5:7b: an actual call through the application adapter returned three valid exact-excerpt mappings. The model has no tools, acceptance authority or payment credentials. Source-excerpt validation happens outside the model.
The PayPal integration uses the actual sandbox Invoicing v2 API for draft creation, retrieval and reconciliation. Evidence, criteria and invoice previews are fingerprinted. SQLite reserves an attempt before the network call. An uncertain result requires reconciliation rather than an automatic create retry. The application has no invoice-send, capture or production-payment endpoint.
AI assisted the project's research, implementation, testing, design and documentation. Within the product, AI suggests evidence mappings; a person makes every acceptance decision.
Challenges we ran into
An excerpt can accurately quote a delivery note without proving that the claimed work happened. The product therefore distinguishes supplied evidence, AI interpretation and human acceptance. A model cannot run or verify the customer's tests merely by reading a pasted log.
Another challenge was handling an interrupted invoice request. Blindly retrying creation could produce duplicate invoices. We used a durable attempt record, a stable invoice number, explicit uncertain states and provider reconciliation. We also check the actual invoice identity, recipient, line item, total, currency and DRAFT state before treating the draft as verified.
Accomplishments that we're proud of
The working application completed a real AI review and created, retrieved and reconciled a real PayPal sandbox draft. Repeating creation returned the same invoice. No invoice was sent and no payment was collected.
The demonstration uses a clearly fictional project, saved actual AI mappings and recorded human decisions. During the film, the app retrieves and verifies the existing draft again without creating another invoice. It also shows the genuine comparison grid, exact source excerpts and an unsaved review of a separate unaccepted milestone whose invoice remains locked. Designed opening and closing graphics frame the actual walkthrough, with natural neural narration and guided close-ups.
Automated workflow checks cover human acceptance, changed evidence, concurrent requests and uncertain responses. Browser checks exercised both themes, keyboard operation and narrow phone layouts; automated accessibility scans found no violations in the tested states.
What we learned
AI is helpful when it makes a review easier to inspect, especially when every claim points back to the supplied source. It should not collapse a nuanced delivery decision into an unexplained score or silently turn an interpretation into a bill.
Building the workflow strengthened my interest in backend systems: persistence, validation and reconciliation determine whether a polished interface can be trusted. The experience also made the practical career value of AI clearer—understanding how to integrate it responsibly is as valuable as generating an answer.
What's next for Milestone Desk
The next validation is a supervised review with freelancers using non-sensitive, consented project examples. We would study whether the evidence view makes disputed or incomplete criteria easier to notice before an invoice is prepared. No customer adoption, time saving or reduction in disputes has been established.
The current product is a private single-owner application with fictional sample data and unsent sandbox drafts. Shared client review, authentication, document integrations, retention controls and a carefully authorized send workflow are future work. A draft is not a sent invoice or a payment, and the application does not make legal or accounting determinations.
Built With
- ag-grid-community
- css
- html
- javascript
- node.js
- ollama
- openai-compatible-http-api
- paypal-invoicing-v2
- paypal-oauth-2.0
- qwen2.5
- sqlite
