$ ls projects
Systems, not stacks. Each entry is a real workflow encoded into agents, constraints, and a human-in-the-loop checkpoint — described by how it behaves in production.
Agentic website generation & optimization platform
flagshipAt a travel/booking software company · for tour & activity operators
I'm building a system that transforms an operator's existing website into a branded, CMS-managed, production-ready site. The interesting part isn't "generating a website" — it's encoding a high-quality web-agency workflow into a repeatable production system that agents can execute safely.
// The staged workflow
- ▸ Crawl & archive the source site as ground truth.
- ▸ Structured business & brand analysis — extract what the operator actually is.
- ▸ Design-system generation, then CMS-bounded implementation.
- ▸ Schema validation with repair loops — malformed output feeds back, never ships silently.
- ▸ Visual QA via Playwright/browser inspection, screenshots, and diffs.
- ▸ Human review at the checkpoints that matter, then CI/CD deployment.
- ▸ Post-launch optimization — monitoring and improvement after go-live.
// Why it's hard (and what makes it safe)
Production-grade agentic systems are mostly the harness, not the prompt: constrained tools, reusable components, deterministic schemas, repair loops, browser-based validation with screenshot evidence, Git-based rollback, and human-in-the-loop checkpoints. That's the difference between a demo and something I'd put in front of a customer — and the thesis behind most of what Iwrite.
// The output
A real production website artifact: Astro + Tina CMS, reusable sections and components, structured content collections, static-site generation.
// Where it's going
Continuous optimization: monitoring, performance checks, SEO/conversion improvements, A/B testing, analytics review, and safe regeneration that improves the site without losing the existing production version. Designed to support scale.
caro.sh
personal productTerminal mastery, unlocked
A product I build and ship solo — focused on helping developers get genuinely fast at the terminal/CLI. It's where I practice the craft I write about: shipping a real product end to end. caro.sh
// The category I build for
A growing market needs senior IC / principal engineers who can enter ambiguous enterprise workflows, understand the business process, and build production-grade AI systems around it: agents, LLM orchestration, MCP/tool layers, RAG-style context acquisition, validation, human review, and cloud-native delivery. That's the forward-deployed / agentic-systems builder lane — the roles emerging at frontier AI labs (Anthropic Research Tools, OpenAI Forward-Deployed Engineer) and principal/ greenfield builder roles in the enterprise.
Adjacent work I draw on: MCP-based and SDK-oriented integrations for AI/search tooling and developer platforms, and hands-on agent-workflow practice across Claude Code, Codex, and Cursor — skills/rules, repository structure for agents, agent handoff, and validation loops.