✦ Writing8 posts

Notes from the engine room.

scribbled between deploys

Practical write-ups on AI engineering, event-driven systems, data pipelines and shipping to production: the trade-offs, the failure modes and the code.

fresh off the press

RAG vs RAG-less: when long context beats a vector pipeline (and when it doesn't)

Million-token context windows and agentic search changed the retrieval trade-offs. A practical framework for choosing between classic RAG, long-context prompting and agentic retrieval.

AI EngineeringRAGLLMs

Sep 30, 2026 6 min read

KafkaEvent-Driven

Event-driven architecture with Kafka: the parts that actually bite you

Topics, partitions, consumer groups, ordering, idempotency and dead-letter queues: the Kafka concepts that decide whether an event-driven system stays correct under load.

Sep 30, 2026 6 min read

Data EngineeringKubernetes

Large data migrations on Kubernetes with Argo Workflows

Modelling a migration as a DAG: chunking, retries with backoff, checkpoints, idempotent writes, backfills and verification, all running as Argo Workflows on Kubernetes.

Sep 30, 2026 5 min read

AI EngineeringVoice AI

Building a low-latency voice AI pipeline: STT → LLM → TTS

Where the milliseconds go in a voice agent, and how streaming, voice activity detection, barge-in and a latency budget make a conversation feel natural.

Sep 30, 2026 5 min read

SecurityAuth

RBAC done right: roles, permissions, JWTs and least privilege

Designing role-based access control that survives growth: permissions over roles, short-lived JWTs with refresh-token rotation, and mapping app roles to cloud IAM.

Sep 30, 2026 5 min read

RedisPerformance

Caching with Redis: patterns, TTLs and the stampede problem

Cache-aside, write-through and TTL strategy, plus the failure modes that hurt in production: stampedes, stale data and invalidation, with code.

Sep 30, 2026 5 min read

DevOpsKubernetes

Zero-downtime deploys: from a single VM to Kubernetes

Rolling updates, blue-green and canary releases, readiness probes and graceful shutdown on Kubernetes, and how to get the same guarantees on one VM with PM2 and Nginx.

Sep 30, 2026 5 min read

AI EngineeringLLMs

Evaluating LLM features in production without fooling yourself

Golden sets, offline evals, LLM-as-judge and its pitfalls, and the online signals that tell you whether an AI feature actually got better.

Sep 30, 2026 5 min read