Saraf

Viewing the market runs top to bottom through the stack: a MarketViewModel, resolved through the Service Locator, subscribes to PricePollingService's Combine publisher, which itself runs a ten-second Task loop calling GetListedPricesUseCase (guarded by an NSLock, paused and resumed across foreground/background transitions) through MarketRepository, through NetworkService, against a typed APIEndpoint. Tapping an asset routes into the PDP (Product Detail Page) feature; tapping Buy there runs its own quote-refresh timer and validates the entered amount before the primary action enables. If the account isn't yet L1-verified, the flow redirects into an auth-pending or auth-declined sheet instead of a trade confirmation.

The parts that fought back.

A region-gated real-time channel, straddling two auth models

SocketIOService.connect() only opens its socket when the device timezone is Asia/Tehran, and it still reads its session token from a legacy global static rather than the new DI-managed auth state. It's a real seam between the old auth model and the new one, and it isn't closed yet.

Backend-specific error mapping, one key at a time

NetworkService handles HTTP 400 through 5xx individually and decodes a shared error DTO carrying both a localized message and an error key. Call sites like KYCL1UseCase switch on that key to surface a specific mismatch (a code/birthdate error reads differently to the user than a code/phone error) instead of one generic failure.

Correctness of a Task-based polling loop across app lifecycle

PricePollingService guards start() and stop() with an NSLock and explicitly listens for foreground and background notifications to cancel and relaunch its polling Task. A naive Task-loop poller's real failure mode is orphaned loops or duplicate timers across suspend and resume, and this is what avoids it.

Coexistence bugs, surfacing mid-migration

During the international/guest-currency rollout, guest and authenticated state weren't threaded consistently through the newer Coordinator navigation. A missing argument in PDP navigation and an unhidden sell button for guests both shipped, and both were caught and fixed the same day.

Identity and money, checked server-side first.

Saraf is a trading app tied to real KYC documents and bank accounts. The security model centers on who's allowed to trade and how session state is handled, more than on cryptography for its own sake:

• Auth and session → Keychain. A dedicated Keychain wrapper, with its own unit tests, replaced the pattern of keeping a session token in a global static string used directly in HTTP headers. That pattern still exists in a few unmigrated legacy code paths.
• OTP → backend-driven, not client-timer-only. Dedicated delivery-status and verification-result entities, plus a resend flow, reflect the backend's actual OTP delivery status rather than trusting a client-side countdown.
• KYC (L1) → server-side identity check, client-side format validation only. KYCL1UseCase sends national ID and birth date to the backend for matching (no identity verification happens on-device), while legal-name fields are restricted to Persian/Arabic script, with digits rejected before the request is ever sent. Trade access gates on L1 status.
• Transport. REST over HTTPS, through an ephemeral URLSessionConfiguration with no on-disk URL cache, reducing residue of API responses left on the device.

Because this is a live trading app, deeper details on token handling and the legacy card/address-verification flow are intentionally left out here.

What I'd tell myself at the start.

• Migrating architecture underneath a shipping app works as a strangler-fig, not a freeze-and-rewrite. The legacy tree still compiles 600-plus commits into the effort, and that's fine: it's being replaced on a schedule, not overnight.
• A documented DI system pays for itself once module count passes roughly a dozen. DI_DOCUMENTATION.md is the one part of the new architecture with its own written usage guide and dedicated performance-test suite, worth the up-front cost at 25 modules.
• Bolting a second product surface onto an in-progress migration multiplies the seams that break. The guest/Arabic-market rollout produced same-day fixes because the Coordinator-navigation contract wasn't fully guest-aware yet when that work started.

A migration in its final stretch, not a finished one.

Saraf's repo shows continuous activity through mid-2026, with June as the busiest month in the project's history. A UI-Finished tag lands in June, and the app's version currently reads as a from-scratch 1.0, consistent with a relaunch rather than an incrementing release. The legacy tree is still there, still compiling, and still shrinking one migrated flow at a time.

The work runs in parallel with BankID: I was promoted within the first year on Saraf and received an internal "Outstanding Performance" recognition for moving a live, KYC-gated trading app onto a modular architecture without a rewrite freeze.