I'm Oleksandr Derechei — I take AI agents from prototype to production, and build the Claude Code plugins, multi-agent systems, and cloud architecture that get them there.
Open-source agent tooling, AI apps, and guides for the Software 3.0 era.
Turns a feature brief into a Well-Architected AWS system — requirements, architecture, IaC, cost, security, observability. A 5-agent pipeline with 6 MCP servers and 34 validation gates that grounds every decision in live AWS docs, so the output is production-defensible, not hallucinated.
94.8% of top sites fail basic accessibility. This drops an AI layer onto any page at runtime — six toggles (AI alt text, plain language, audio, captions, translation, AAA contrast) plus a screen-reader simulation — taking a failing page to WCAG AAA with zero redesign. Claude Vision + Web Speech, live demo.
Audits any codebase in any language for God Classes, Anemic Models, and Feature Envy, then ships gated, test-verified refactors toward GoF patterns — never touching code that fails your own tests. Principle-based detection over a 102-entity glossary (45 smells + 57 patterns).
A fast, themeable statusline that surfaces git, tokens, context window, and rate limits in Claude Code at a glance — 38 widgets, a live TUI editor, zero network on the hot path. Shipped on npm and installed with one command.
One article, every audience: AI reshapes the same content by age, language, gender, and format on demand, so writing meets each reader where they are. A Svelte 5 + Claude API product with a live demo.
Turns Karpathy's LLM reading list into an operable, agent-driven knowledge base in Obsidian — a four-layer, hook-enforced agent stack with multi-agent orchestration that keeps every generated page provenance-tracked and current.
Treats your AI coding setup as a system to be tuned: an evaluate → plan → calibrate → re-evaluate loop that finds what's misconfigured, fixes it, then scaffolds enforcement — a hook, rule, or wrapper skill — so the same failure can't recur.
The home of the whole stack — a Claude Code marketplace distributing every plugin I ship under @odere-pro. Add it once, install any tool individually, and browse the catalog with live search. One ecosystem, versioned and maintained.
The playbook behind the rest of this work: an 8-part field guide for taking teams from ad-hoc AI prompting to a disciplined, production-grade agentic workflow — rules, context, verification, cost control, team standards. Inspired by Karpathy's Software 3.0; read in 11 languages.
From prototype to production.
I'm a forward deployed AI engineer — I embed with the problem, take agents from prototype to production, and build the plugins, multi-agent systems, and cloud infrastructure that keep them reliable at scale. The primary user is increasingly an LLM, so I design for it: stable contracts, prose-for-agents docs, and deterministic gates that hold whether a human or an agent is driving.
I ship in the open. The odere-pro stack spans a published npm CLI and a Claude Code plugin marketplace — claude-aws-architect alone runs 5 agents, 6 MCP servers, and 34 validation gates, grounding every claim against AWS docs. As a founder, I turn agent-native tooling into products teams depend on — moving fast without trading away reliability.