Weekly learnings
What I learned this week
A short note most weeks — small discoveries, fixes and ideas from day-to-day building. The longer stories live on the blog.
- 11 Oct 2026 Test-time compute, make it work first, and Claude Code with Jira A week covering what test-time compute means, a build-order rule learnt the hard way, and using Claude Code with Jira to keep a backlog clean.
- 04 Oct 2026 The revenue cost of latency at Amazon scale A note on how every extra 100ms of latency cost Amazon about 1% of sales, and what that would mean in revenue today.
- 27 Sept 2026 Training an LLM from scratch, and why retrieval beats a better model in coding agents A hands-on LLM-from-scratch tutorial and a Microsoft fastcontext stat suggesting smarter retrieval, not a bigger model, is the biggest lever for coding agent efficiency.
- 20 Sept 2026 Keycloak, Claude Code's headless OAuth tokens and language model types Exploring Keycloak as an Entra ID alternative, running Claude Code headless with a subscription OAuth token, masked versus autoregressive language models, and why evals matter more than demos when building AI products.
- 13 Sept 2026 When a script beats an agent, the N+1 query problem and database transactions Why not everything needs to be an agent, Conventional Commits as an org-wide standard, the classic N+1 query problem, wrapping writes in transactions and Archify for visualising codebases.
- 06 Sept 2026 Graphify's token trick, Autocannon load testing and a system design tutorial How Graphify cuts token usage by moving search out of the model, Autocannon for CLI load testing, favourite VS Code themes and a practical system design tutorial.
- 30 Aug 2026 LLM gateways, guardrail latency and quality checks in CI Cloudflare Vectorize for the site chatbot, Lighthouse and Lychee as CI quality gates, and AI application lessons on guardrail latency, prompt-injection defenses, gateways, caching risks and chatbot feedback.
- 23 Aug 2026 Astro, self-publishing automations and enterprise knowledge bots Building this site with Astro, automating weekly publishing with Claude Code, brainstorming an enterprise search bot, picking Sentry for error tracing, and queuing up my next read.
- 27 Jul 2025 MCP servers and production-grade AI agents Wiring MCP servers into Claude and Cursor, and a blueprint for taking AI agents from dev to production.
- 20 Jul 2025 FTI architecture, Uber's QueryGPT, Qdrant and a diagram trick Structuring LLM apps with Feature/Training/Inference pipelines, natural-language-to-SQL at Uber scale, a free vector DB for demos, and mermaid → draw.io.
- 06 Jul 2025 A system design goldmine One find this week: a huge curated collection of system design resources.
- 29 Jun 2025 Graph databases, dbt hands-on, AWS Summit — and starting this notes site A packed week: Neo4j vs DGraph with LangChain, a full dbt walkthrough, notes from AWS Summit India, and the first version of this weekly-notes habit.
- 22 Jun 2025 AWS container services, and tracking LLM benchmarks Comparing ECS, EKS, App Runner and Lambda from a GCP/Azure background, plus a great site for model comparisons.