Creative by nature.

Customers, providers, and administrators need different tools, but they’re working on the same service. I connected the booking, fulfillment, conversation, pricing, and operations workflows around shared records.

Conceptual architecture based on the TLC apps and shared Supabase order/pricing workflows. The website adds discovery, signup, scheduling, and support. This diagram simplifies the system; it is not a claim that every client receives every update instantly.

I built the checks around payment and payout actions: saved amounts, current permissions, valid financial states, and recovery when a response is uncertain. A retry should start by finding out what already happened.

The September payment work recorded 32 mocked handler tests covering permission, stale writes, failed saves, authorization/capture errors, and retries. October Grow finance work recorded 18 backend tests for server-owned reviews, expiry, duplicate confirmation, destination changes, and unknown outcomes. Backend deployment is documented; these checks did not send a real instant payout or charge a live card.

A hidden button is not an access rule. For order history and proof photos, I restricted access to the order’s customer, eligible assigned provider, or an administrator—and checked unrelated and signed-out cases.

October 8 deployed database checks verified participant/admin reads and unrelated/signed-out denial for the scoped history endpoint and proof-photo policies. All existing proof objects and status events were retained. Actual signed-JWT file transfer was a separate acceptance step. These are specific implemented controls, not a whole-system security certification.

I connected Stripe’s current prices to the shared catalog used by the apps and website. New orders use updated rates; accepted orders keep their saved prices. Invalid catalog changes stop before publication.

October 15–16 records cover service-only publication, single-use claims, duplicate/invalid price rejection, atomic updates, and configuration conflicts. All 44 existing orders retained prior item prices and totals; six unfinished orders gained only missing optional items. A new optional item does not itself charge a customer.

I connected the assistant to approved guidance and current pricing, with narrowly scoped profile context. If facts are unavailable or the request needs detailed account access, it gives an honest fallback or a path to human help.

Consumer Ask TLC launch scope is guidance plus the caller’s own permitted profile facts. Detailed orders, jobs, and payments are excluded; Grow’s admin assistant has a separate scope. October launch/pricing regressions cover authorization and unavailable pricing. Signed-in end-to-end acceptance is distinct from backend tests.

Order activity belongs beside customer and provider messages. I projected recorded status events into the conversation without inflating unread counts, and separately improved iOS push routing and credential reuse.

The October status reader preserves recorded timestamps, checks order membership, supports paging, inserts no chat records, and sends no extra push. The September iOS notification repair reused APNs credentials within a worker and shared concurrent refreshes. A targeted sandbox push was accepted by Apple; actual phone receipt remained a separate check.