Principles
Most of what makes software last is decided before the first line of code. We write that part down, then automate everything after it.
- Specs before code: APIs, events and screens are agreed in writing first.
- One repository and one pipeline for clients, services and infrastructure.
- A reproducible development environment: a new laptop runs the whole system with one command.
- Every change is reviewed, infrastructure included.
- Boring technology where it is enough.
Mobile
The apps are built on Kotlin Multiplatform: the domain, data and networking are written once and compiled natively for iOS and Android. The UI is chosen per product — SwiftUI and Jetpack Compose where each app should feel at home on its platform, Compose Multiplatform where one screen can serve both.
- Shared code compiles to native binaries — no virtual machine, no JavaScript bridge.
- Shared logic covered by tests that run on the host and on both platforms.
- Offline first: the app works without a network and syncs when it can.
- Snapshot tests for every screen in light and dark themes and at large text sizes.
Backend
Services in Go and Kotlin, generated from GraphQL, OpenAPI and protobuf specs: the spec is the contract and the code follows it. Events go through Kafka; long-running processes run on Temporal.
- Server and client code generated from the same specs.
- Kafka with strict ordering, idempotent consumers and a dead-letter queue.
- Temporal for processes that take minutes or days: payments, onboarding, retries.
// Every event carries an ID. Seeing it twice is a no-op.func (c *Consumer) Handle(ctx context.Context, m kafka.Message) error {
id := eventID(m)
return c.db.InTx(ctx, func(tx Tx) error {
done, err := tx.Processed(ctx, id)
if err != nil {
return err}
if done {return nil
}
if err := c.apply(ctx, tx, m.Value); err != nil {
return err // retried; after N attempts, sent to the DLQ
}
return tx.MarkProcessed(ctx, id)})
}
Infrastructure
Kubernetes, with Terraform / OpenTofu for everything around it, including DNS and alerts. If it isn’t in the repository, it doesn’t exist.
Clients
- iOSKotlin · Swift
- AndroidKotlin
Services
- API gatewayGraphQL · OpenAPI
- Domain servicesGo · Kotlin
Data
- PostgreSQLstate
- Kafkaevents
- Environments created from code, identical except for size.
- Infrastructure changes go through the same review as application code.
- Secrets live in a secret store, never in the repository or the app.
Observability
Every request leaves a trace. We alert on what users feel, so the pager rings for real problems, not for a busy CPU.
- OpenTelemetry traces across clients and services.
- Structured logs, linked to traces.
- Metrics and dashboards for every service and endpoint.
- Alerts on errors and latency, not on CPU.
Delivery
A merge to main is a release candidate. The pipeline does the rest; people decide when it goes out.
- Lint and tests, in parallel for every platform and service.
- Build and sign; the version comes from the commit.
- Internal tracks: TestFlight, Google Play internal testing, staging.
- Staged rollout to production.
AI in the loop
Coding agents work in our repositories with project skills and MCP servers that know our specs, templates and tools. Agents write tests, scaffold services and do the first pass of review. Decisions are made by people.
- Project skills: our conventions, written down for agents.
- MCP servers for specs, CI, logs and tickets.
- Agent changes go through the same review and pipeline as everyone else’s.
Problems worth a note
Strict ordering without losing throughput
Events for one account must be handled in order; events for different accounts must not wait for each other. Kafka partitions by key give the order, and idempotent consumers make retries safe, so a partition never stalls on a duplicate.
Real time without blocking
In a voice pipeline, audio sometimes arrives faster than it can be processed. Go channels with a non-blocking try-send drop a frame instead of stalling the stream, so latency stays flat under load.
Need something like this built?
We take products from the first screen to the store release, and keep them running after.