Kubai Kevin

Older Posts — Page 6 of 19

Platform engineering ROI: metrics that fund teams

How to connect platform work to business outcomes with flow metrics, honest attribution, and instrumentation you can build in an afternoon.

13 min read · 19 September 2026

Context windows: agents that run for days

Long-running agents fail when context is treated as a log. Keep constraints verbatim, checkpoint durable state, and cap the working set per cycle.

17 min read · 19 September 2026

Most LLM cost dashboards lie…

Per-request token attribution, prompt caching, model routing and retry caps: how to instrument LLM spend so the numbers match the invoice.

13 min read · 19 September 2026

Mobile money agents: when M-Pesa API fails

Mobile money APIs return pending, then settle asynchronously. Learn to poll, deduplicate callbacks, and reconcile orphaned transactions.

14 min read · 18 September 2026

Spot instances vs on-demand for AI experiments

How spot and on-demand GPU instances differ on interruption handling, effective throughput, and the spend threshold where spot engineering pays off.

13 min read · 18 September 2026

Local LLM agents: vLLM vs Ollama for fintech

Compare vLLM and Ollama for on-prem LLM agents in regulated fintech: concurrency, guided JSON, VRAM budgets, cost math, and a 30-minute test.

16 min read · 18 September 2026

Multi-agent systems: why failures hide

Why multi-agent LLM pipelines fail silently, and how per-agent tracing, schema validation, budget enforcement, and loop detection make those failures loud.

13 min read · 18 September 2026

Agent drift: the failure mode you didn't log

Agent drift is a distribution failure, not a model failure. How to detect it with categorical signatures and PSI, and contain it with state guards.

15 min read · 18 September 2026

Feature flags became AI rollout platforms

How flag platforms handle structured model configs, eval-metric gating, and offline fallback — plus a checklist for choosing one for AI rollouts.

18 min read · 18 September 2026

Agent access for non-AI engineers

A backend-first guide to exposing internal tools to LLM agents: typed schemas, dry-run defaults, idempotency, budgets, and replayable traces.

16 min read · 17 September 2026

Stop Putting LLM Keys in .env

Why flat .env files fail at scale for LLM API keys, and how a secrets manager plus gateway brokering scoped credentials fixes rotation and leaks.

14 min read · 17 September 2026

Which agent stack survives 3G?

How edge-deployed and centralised agent stacks behave on high-latency, lossy mobile links, and how to measure reconnection storms before choosing one.

13 min read · 16 September 2026

Self-service AI tooling: where it breaks first

Self-service AI platforms fail on shared keys, missing cost ceilings and model drift. How to spot each pattern, instrument it, and fix it at the gateway.

15 min read · 16 September 2026

Agent platform abstractions: a case study

Why agent failures are platform failures: typed tool contracts, durable execution, and structured traces that make runs replayable and retries safe.

11 min read · 16 September 2026

Zero-trust AI: skip the latency tax

How to keep per-request identity and least privilege in inference services while moving handshakes and policy evaluation off the hot path.

14 min read · 15 September 2026

Memory Systems for Production Agents

How in-memory stores, caches, and stateful services leak context in production agents, plus code, failure modes, and instrumentation to measure it yourself.

11 min read · 15 September 2026

Stop trusting agents for core logic

Agent orchestration hides coupling in YAML and makes review harder. Learn contract-first design, policy-as-code, and how to measure the trade-off.

11 min read · 14 September 2026

Agent API versioning: why clients break on upgrade

Why agent version upgrades break downstream clients, the contract drift behind the errors, and how to pin, check and roll out versions safely.

12 min read · 14 September 2026

Designing a Schema That Survives Embedding Model Upgrades

Why embedding columns break on model upgrades, how to version vectors, migrate legacy JSON payloads, and when to offload to a vector store.

11 min read · 13 September 2026

Stop assuming payment APIs never fail

How to combine AI recommendations with M-Pesa, Paystack and Flutterwave calls so gateway timeouts, rate limits and duplicate charges don't break checkout.

13 min read · 13 September 2026

Platform Teams: Agentic World Struggle

Platform teams must adapt identity, tooling, and observability for AI agents that call APIs instead of using dashboards, with concrete patterns and failure modes.

13 min read · 12 September 2026

MCP hidden cost traps

How misconfigured autoscaling, over-permissive IAM and secret-lease churn create a cost feedback loop on managed container platforms, and how to break it.

13 min read · 12 September 2026

Agent Systems: Circuit Breakers Are Not Enough

Circuit breakers and bulkheads were built for stateless services. Agent systems need semantic error handling, externalized state, and recursion guards.

13 min read · 12 September 2026

Why ML Models Drift After Deployment

Offline validation rarely matches production traffic. Learn the evaluation gap, how to measure drift, and concrete checks to add before your next deploy.

16 min read · 11 September 2026