Writings on agentic AI, LLM engineering, and software development.
I turned a free Oracle Cloud VM into a private DNS and VPN layer using Tailscale and AdGuard Home — and learned that the hardest part of security architecture is deciding what not to expose.
MCP solved agent-to-tool. A2A solved agent-to-agent. The layer nobody's built yet is memory-to-memory — and three separate research directions (Mesh Memory Protocol, multi-agent computer architecture, Portable Agent Memory) are converging on the same missing primitive: consistency, not retrieval.
Most habit apps make you feel bad when you slip. Rite tracks your daily rituals, uncovers patterns you can’t see yourself, and explains them plainly—without judgment, guilt trips, or ads. Here is why we built Rite differently.
Stop acting as a glorified onboarding manager for an amnesiac assistant. Here's how CLAUDE.md ends the re-explanation tax, what it actually costs to maintain, and the exact rule I use to keep mine lean.
I stopped forcing agentic work into sprint boards and standups. Here's the loop that quietly replaced them — and why "Verify" is now the load-bearing wall of my entire process, not a nice-to-have step.
I went four days without writing a function by hand — and it was some of the most technical work I'd done in months. Here's what actually fills my day now that "coding" doesn't mean typing anymore.
Everyone calls themselves an "agentic engineer" now. Most are still vibe coding with extra steps. Here's the actual line between the two — and why I had to learn it the hard way rebuilding SYSLAB twice.
I rebuilt a single overloaded AI agent into a swarm of specialists and watched it get more resilient — and smarter. Here's why I think decomposed, cooperating agents beat one model trying to hold an entire problem in its head, and where the orchestration pain is real.
I ran one deliberately overloaded prompt through 6 AI image outputs and judged them blind on prompt adherence, anatomy, lighting, and storytelling. The winner wasn't the prettiest — it was the most coherent. Here's what actually separated a 79 from a 96.
AI coding tools let you ship full-stack apps overnight — but speed kills security. From JWTs in localStorage to wildcard CORS and exposed databases, vibe-coded apps are riddled with exploits. Here's a 31-point security checklist to audit and harden everything before it costs you.
Traditional system design learning relies heavily on articles and videos. Discover how simulation-based learning helps engineers understand scalability, reliability, and architecture trade-offs through hands-on experience.
Why I stopped building a multi-tenant AI agent platform and started building a four-agent shadow organization for RiteLabs instead — the philosophy, the rejected designs, and the architecture that survived four rounds of review.