Client work.

Traders depositing in multiple currencies were funnelled to a single payment provider, so deposits failed whenever that provider had localised outages or weak success rates in specific corridors — stalling onboarding in regulated regions.

A routing layer that selects a PSP per request on currency, region, KYC tier and live success rates, with per-provider adapters behind one internal contract, idempotent request handling, and a fallback chain for failed authorisations.

One provider's outage no longer fails a deposit — an affected corridor falls back to another PSP on the same request, instead of returning the trader an error. Onboarding a new provider dropped from weeks to days.

Retries from upstream PSPs and asynchronous webhook deliveries occasionally produced duplicate postings against trader wallets, forcing finance into manual clawbacks and exposing the firm to regulator-flagged clawback risk.

A deposit and withdrawal pipeline rebuilt around deterministic idempotency keys persisted in PostgreSQL, with an explicit state machine per in-flight transaction, replay-safe webhook handlers, and workers that retry to a terminal state — plus a dead-letter queue and operator tooling for stuck entries.

Eliminated double-postings in production. Nightly reconciliation moved from manual review to an automated daily report with exceptions only.

Reconciling internal wallet ledgers against PSP statements was a manual end-of-day exercise that could not keep up with trade volume; discrepancies surfaced hours or days after the fact.

A pipeline that ingests PSP statements (CSV, API, SFTP-delivered) into a normalised event store and matches them against the internal ledger on deterministic and fuzzy rules, surfacing unmatched entries in an operations UI with one-click resolution.

Daily reconciliation dropped from a multi-hour manual review to an automated run with a small exception queue. Discrepancies are now caught the same day they occur.

Internal services exposed inconsistent authentication schemes and rate limits, blocking partner onboarding and pushing security policy into every team's backlog.

A gateway across .NET Core, Node.js and Azure API Management fronting all partner-facing services with unified OAuth2, throttling, request shaping and routing — with services migrated behind it incrementally, without disrupting active integrations.

Partner integration time dropped from weeks to days. Security policy is now enforced in one place rather than audited per service.

A monolithic legacy product could not be extended to support multi-tenant deployments and hierarchical project-group structures without taking the whole system down for every release.

A distributed architecture designed from scratch around domain-driven service boundaries, with structured project-lifecycle workflows, role-based access over hierarchical data, clean service contracts and message-driven coordination.

Each service ships independently, enabling continuous delivery for the first time. New tenants onboard without code changes and team-level rollouts no longer block product-wide releases.

As trader volume grew, several backend services were hitting tail-latency budgets — slow synchronous database queries and chatty service-to-service calls dominated the hot path during market opens.

An end-to-end profile of the hot path, then Redis caching for read-heavy reference data, non-critical work moved onto RabbitMQ workers, the affected services containerised with sized resource limits, EF Core query patterns tuned, and per-endpoint latency budgets enforced in CI.

The endpoints that were breaching their tail-latency budgets now hold them through market open — and each budget is enforced per endpoint in CI, so a regression fails the build instead of reaching traders.

Every new client deployment required engineers to provision and configure servers by hand — slow, error-prone and inconsistent. Environment drift led to reproduction-only bugs in production.

A provisioning toolchain: an Angular operator UI, a .NET Core Web API as the control plane, and Ansible playbooks doing the infrastructure work — every deployment idempotent and parameter-driven.

New client environments now stand up in minutes instead of days. Environment drift between clients and across release cycles has effectively been eliminated.