Articles
Engineering writing on AI agents, retrieval systems, and Angular. Also on Medium.
Orchestrator–workers: when one model should manage others
A supervisor LLM plans at runtime, a deterministic validator catches bad plans before workers run, and dependency waves let independent tasks overlap. When dynamic decomposition earns its overhead and when it doesn't.
July 30, 2026 · LLM Patterns · effective-agents-labParallelization in LLM systems: sectioning and voting
Two distinct reasons to run LLM calls in parallel — sectioning independent subtasks for speed and coverage, voting for reliability — with real cost traces and the concurrency bug that only appears under load.
July 25, 2026 · LLM Patterns · effective-agents-labRouting: the cheapest reliability win in LLM systems
A cheap classifier dispatches to specialized handlers, each with its own prompt and model tier. Low confidence escalates to a human at the cost of exactly one classifier call — nothing more.
July 18, 2026 · LLM Patterns · effective-agents-labPrompt chaining vs one big prompt: when decomposition wins
When does breaking a task into sequential LLM calls beat one big prompt? The case for chaining, the gate that makes it work, and when it just adds latency.
July 17, 2026 · Agents · FundamentalsThe agent loop: the pattern under every AI agent
Strip every framework away and an AI agent is one loop in 81 lines of Python: call the model, check if it wants a tool, dispatch it, feed the result back, repeat. Here's what that loop actually looks like.
July 16, 2026 · Build in public · ProcessDemos are not products: auditing my own GitHub
I went through every public repo and asked one question of each: demo or product? The honest answer was demo, more often than I'd like — and the failure mode was the same in all three cases.
July 15, 2026 · RAG · PatternsThe replay-demo pattern: show your RAG working without an API key
How to demo a RAG pipeline honestly, at zero cost, with no abuse surface — by replaying real recorded benchmark runs instead of calling a model on every visitor request.
July 15, 2026 · RAG · Build logBuilding ClauseLens: what 90.7% retrieval actually took
A hybrid-retrieval contract-intelligence pipeline with zero cloud dependencies — and the three real bugs (preamble-dropping chunker, cross-process persistence, SVD dimension drift) that the benchmark surfaced along the way.