Engineering & Architecture: Build Decisions This Week
2026‑09‑26
The pulse of engineering is still very much a trade‑off exercise. Every decision we make today will be bought or sold against cost, speed, reliability and compliance tomorrow. For CTOs in South Africa, the UK and across Europe this week there are three axes that stand out:
Below is a concise review of what we should be building, why it matters and what can stay on the table for now.
---
Anglo American’s IFN initiative demonstrates that a platform‑centric model can unlock more than just revenue; it can unlock access to R1.8 bn in capital by tying service modularity directly to investor upside【[1]】. The implication for our stacks is straightforward:
| Decision | Cost | Benefit |
|----------|------|---------|
| Adopt a service‑mesh + API gateway pattern, exposing core capabilities (identity, billing, analytics) as reusable micro‑services | Initial refactor, onboarding overhead for teams | Enables rapid composability; external partners can consume APIs on demand, lowering friction for new funding rounds |
| Keep legacy business logic in an isolated “monolith‑to‑micro” hybrid until critical revenue paths are stabilized | Potentially higher short‑term complexity | Protects mission‑critical workloads while giving investors a clear, auditable revenue engine |
What to ignore this week – pushing all code into the cloud at once. In South Africa the bandwidth variance can still be significant; an incremental migration protects against over‑provisioning and ensures that any downstream latency problems are caught early.
---
Gergely Orosz’s interview with Maggie Appleton in Pragmatic Engineer highlights how internal AI agents can surface patterns, enforce standards and reduce friction during code reviews【[2]】. Yet it also stresses that “human judgment still matters.”
| Decision | Cost | Benefit |
|----------|------|---------|
| Integrate an open‑source LLM (e.g., turbopuffer) into the CI pipeline for static analysis, comment suggestion and documentation generation | Compute cost on a per‑run basis; model fine‑tuning overhead | Lower review turnaround times; consistent style enforcement |
| Maintain manual “trust‑but‑verify” checkpoints for security‑critical modules | Extra reviewer time | Protects against false positives/negatives in AI outputs |
What to ignore this week – building a custom LLM from scratch. The incremental approach of wrapping an existing model in the workflow keeps you within budget while still reaping early wins.
---
The FX market is shifting again: “Old‑school voice brokers face automation in the $4 trn FX market”【[5]】. This transition isn’t just about speed; it’s also a test of resilience and compliance. For a platform that may expose financial data to multiple jurisdictions, consider:
| Decision | Cost | Benefit |
|----------|------|---------|
| Deploy an algorithmic matching engine (e.g., FIX‑based) behind a low‑latency queue in the cloud, with multi‑region failover | Network and vendor licensing costs; developer time for integration | Handles 10× more volume with sub‑millisecond latency |
| Keep a legacy voice channel as a “fallback” only during migration | Parallel maintenance cost | Reduces risk of service disruption during transition |
What to ignore this week – full replacement of the voice system overnight. A phased hybrid model protects revenue streams while still moving toward the scalable core.
---
| Focus Area | South Africa (POPIA) | UK/EU (GDPR, AI Act) |
|------------|---------------------|----------------------|
| Data Residency | Prefer local data centres in the Cape/Western Cape to satisfy POPIA’s “data location” clause | Use EU‑centric regions for European users; ensure data is not routed through non‑EEA locations |
| Auditability | Build immutable logs (write‑once) into every API gateway to satisfy both POPIA and potential future audit regimes | Implement dynamic pricing logs in compliance with UK class action precedent; log all price calculations so the platform can “prove” fairness |
| Security Posture | Use AWS Shield / Azure DDoS Protection for public APIs; enforce TLS 1.3 + HSTS | Apply EU AI Act guidance on explainable AI outputs in automated decision systems |
---
These actions align with the capital injection spotlighted by Anglo American, respect the nuanced balance of AI and human judgment highlighted by Maggie Appleton, and address the performance demands emerging from automation in FX trading.
---
---
The above build decisions assume that:
Please validate these assumptions against the engineering team’s capacity, budget constraints, and any region‑specific regulatory reviews (POPIA, GDPR). The trade‑off between immediate speed to market and long‑term maintainability should be quantified in a risk register before committing to full migration.