Engineering & Architecture: Build Decisions This Week
2026‑11‑20
The week has been a whirlwind of AI‑centric headlines and a surprisingly quiet political moment in UK football. While the Premier League clubs’ vote to cover the legal costs of their directors over a Manchester City sponsorship dispute (as reported by City AM in “Premier League clubs back directors over Manchester City legal threat”) is unlikely to influence our day‑to‑day stack, it underscores a broader truth: institutional decisions can hinge on risk‑management and governance, the same principles we apply to engineering choices.
Below are three concrete build decisions that align with the latest industry thinking from the Pragmatic Engineer and ByteByteGo newsletters. These actions are applicable in both South African and European environments, where cost, bandwidth, and talent availability continue to shape architectural trade‑offs.
---
Inference engineering for software engineers (Pragmatic Engineer) details how Uber, Pinterest, and AT&T have slashed per‑token costs by 50 %–90 % by shifting from frontier‑lab APIs to open models. The article presents the classic “bare‑metal vs. cloud GPU” decision matrix:
| Option | CAPEX | OPEX | Control | Latency | Ecosystem |
|--------|-------|------|---------|---------|-----------|
| Bare‑metal GPUs (on‑prem or colocation) | High upfront, predictable | Low once the hardware is in place | Full control over KV cache and network isolation | Very low, hardware‑bound | Requires in‑house expertise |
| Cloud GPU services (AWS, GCP, Azure) | Zero upfront | Pay‑as‑you‑go, higher for sustained use | Managed, easier scaling | Slightly higher, multi‑tenant | Rapid elasticity, global regions |
Action:
---
In “Software factories and code reviews with Harjot Gill” (Pragmatic Engineer), Harjot Gill explains how CodeRabbit shipped over 5,000 pull requests in a month and chose to invest $250 000 a month in AI tooling, ultimately arguing that human code review is becoming obsolete. The key takeaway is that automation, when coupled with a well‑defined workflow, can surface defects faster and at a lower cost.
Action:
The trade‑offs are clear: higher upfront tooling cost versus reduced human effort and faster delivery. In a talent‑scarce market like South Africa, reducing manual review time can free senior engineers to focus on architectural innovation.
---
ByteByteGo’s “Evals for AI Engineers” newsletter highlights the challenges of moving AI agents to production: data freshness, bounding, and keeping agents within policy constraints. GlobalFoundries’ case study—choosing Redpanda and building a lightweight eval pipeline—illustrates a practical roadmap.
Action:
By integrating these evals early, you reduce the risk of runaway costs or policy violations, a concern that resonates across both UK/EU GDPR and South African POPIA contexts—even if the article does not mention the legislation directly, the principle of bounded access is universally relevant.
---
The Premier League vote may seem unrelated, but it is a reminder that governance matters at every layer—from boardrooms to codebases. By focusing on these three decisions, you align engineering spend with strategic risk mitigation while keeping your team agile and responsive.
---
Sources
Review Note