Engineering

How we build systems that keep working.

A layered approach, explained in plain language. Click a layer to see what it does, how it fails safely and what it takes to secure it. Not every project needs every layer.

Interactive simulation

Architecture explorer

Illustrative reference architecture with fictional data. It does not describe any specific deployment.

Evidence

Engineering in practice

Numbers are only useful with their context. Each figure below says when, where and how it was measured, and what it does not prove. We do not publish uptime or volume figures until they have been measured over a period long enough to mean something.

  • 22.8 s → 14.7 s

    Time until a WhatsApp reply finished sending, production WhatsApp assistant

    Measurement period
    August 2026, before and after one change set
    Environment
    Production
    Sample
    Small sample of real executions
    Method
    Execution timings from the workflow engine, measured to the end of the send step
    Known limitations
    Not a load test. Reply time also depends on the AI provider and the messaging provider on the day.
  • 12 874 → 2 801

    Full-price input tokens per reply after restructuring the prompt cache, same assistant

    Measurement period
    August 2026
    Environment
    Production
    Sample
    11 executions
    Method
    Token usage reported by the AI provider per call
    Known limitations
    Shows lower running cost per reply. It says nothing about conversion or customer satisfaction.
  • Unit tests on every build

    Showroom state machine, contact validation and the contact endpoint (unit, every build); every route and showroom control (browser tests, before each release)

    Measurement period
    From October 2026
    Environment
    Development
    Sample
    Desktop and mobile viewports for browser tests
    Method
    Vitest unit tests run inside the build command; Playwright browser tests run before each release
    Known limitations
    Tests prove the demos behave as designed. They are not evidence about client systems.

Security and trust

Responsible engineering, stated precisely.

We hold no security certifications, and we do not claim that any tool makes a system POPIA compliant. Every statement below is tagged with how firmly we can make it. Publishing this page is not an independent security assessment.

  • Verified on this website
  • Standard practice on our builds
  • Recommended for enterprise builds
  • Agreed per project

Architecture and integration design

  • Standard practice on our builds: Data flows and integrations are drawn and agreed before the build.
  • Standard practice on our builds: Workflow engines and admin tools stay off the public internet, behind a tunnel.
  • Recommended for enterprise builds: A short decision record for each significant architecture choice.
  • Agreed per project: Which systems connect, and who owns each one.

Authentication and authorisation

  • Verified on this website: This website has no logins, accounts or admin area.
  • Standard practice on our builds: Staff consoles sit behind authentication, with the least access each role needs.
  • Recommended for enterprise builds: Single sign-on with the client's identity provider, and periodic access reviews.
  • Agreed per project: Roles, and who may see or change what.

Credential management

  • Verified on this website: Secrets live only in server-side environment variables. A pattern scan of the full code history of this website on 26 October 2026 found none committed.
  • Standard practice on our builds: Credentials are kept in the workflow engine's credential store, never pasted into workflow steps.
  • Recommended for enterprise builds: Separate credentials per environment, with a rotation schedule.
  • Agreed per project: Who holds administrative credentials, and how they are handed over.

Input validation

  • Verified on this website: The contact form is validated on the server, with length limits, control characters removed, a request size limit and an origin check.
  • Standard practice on our builds: Data is validated and normalised at every boundary between systems.

Workflow reliability

  • Standard practice on our builds: Duplicate guards so a retried message or event is not processed twice.
  • Standard practice on our builds: Pause switches that stop automated sending immediately.
  • Recommended for enterprise builds: Retries with back-off and a holding queue for work that keeps failing.
  • Agreed per project: What must happen when each process cannot complete.

Error handling

  • Standard practice on our builds: Workflows are designed with an error path that alerts a person, and silence is treated as a failure.
  • Agreed per project: Who is alerted, by which channel, at what hours.

Logging and monitoring

  • Standard practice on our builds: Execution history in the workflow engine, and health checks after changes.
  • Recommended for enterprise builds: Central logging, alerting and uptime monitoring for production systems.
  • Agreed per project: What is monitored, what is reported, and how long logs are kept.

Human escalation

  • Standard practice on our builds: Urgent, sensitive or irreversible cases go to a person with the full context.
  • Standard practice on our builds: Staff can take over, or hold automated replies for approval before they are sent.
  • Agreed per project: The escalation rules for each process.

Backup and recovery

  • Standard practice on our builds: A snapshot before every change, with a tested rollback.
  • Recommended for enterprise builds: Scheduled backups with restores tested on a copy.
  • Agreed per project: Backup frequency, retention, and how quickly a system must be restored.

Testing and release management

  • Verified on this website: Every build of this website runs the unit tests and fails if it finds inline scripts, source maps or secret patterns.
  • Standard practice on our builds: Browser tests of every page and demonstration are run before releases.
  • Standard practice on our builds: Changes are rehearsed on a copy of live data, then smoke-tested after release.
  • Agreed per project: Acceptance tests, and who approves a release.

Data minimisation and retention

  • Verified on this website: The contact form asks only what is needed to reply, and the website sets no cookies.
  • Standard practice on our builds: Systems collect only what the process needs.
  • Agreed per project: Retention periods, and deletion on request.

Privacy and information security

  • Verified on this website: HTTPS everywhere with HSTS, a strict Content Security Policy, security headers on every page and API response, and a rate limit on the contact form set at the network edge.
  • Standard practice on our builds: POPIA is considered in the design. The client remains the responsible party.
  • Recommended for enterprise builds: An independent security test before go-live for systems that handle sensitive data.
  • Agreed per project: Where data is stored and processed, and the agreements with each provider.

Third-party service dependencies

  • Verified on this website: No third-party scripts, analytics or trackers. Fonts are self-hosted. Production dependencies had no known vulnerabilities at the last audit (26 October 2026).
  • Standard practice on our builds: Every external service a system uses is listed, with what data it receives.
  • Agreed per project: Providers outside South Africa, and cross-border transfers.

POPIA: personal information collected by a system we build is processed for the purpose the client defines, kept to what that purpose needs, and protected with reasonable technical measures. The client remains the responsible party; we help them meet that responsibility in how the system works. Where a control is marked "agreed per project", it is decided and documented for that engagement, not assumed.

Want a technical walk-through?

If you are the person who has to sign off on the architecture, request a confidential technical briefing. We go through the design, the data flows and the failure modes in detail, within what each client's confidentiality allows.