Uses

Operating system & shell

• Windows 11 with WSL2 for development work
• macOS for video and design work when needed
• Terminal: Windows Terminal + zsh in WSL

Code editor

• VS Code with these extensions I actually use:
  • Astro
  • Tailwind CSS IntelliSense
  • GitLens
  • Error Lens
  • Markdown All in One
  • GitHub Copilot (used judiciously - see writing on AI Trailblazer adoption)

Development stack

• Node.js 20 LTS - primary runtime
• Python 3.11+ with FastAPI for AI services
• Docker Desktop + Docker Compose for local infrastructure
• GitHub for version control and collaboration
• PostgreSQL, Redis, MinIO as the standard local services for AI/RAG work

AI / LLM tooling

Daily drivers

• Claude (Anthropic) — primary reasoning and writing model. Best long-context analysis in my hands, and the tool I reach for on complex technical problems, program documentation, and anything requiring careful structured thinking.
• Claude Code — Anthropic’s AI CLI tool for software development. The backbone of how I build and iterate on this site and side projects. Runs in the terminal, has full codebase access, and the MCP tool ecosystem makes it genuinely extensible.
• ChatGPT (GPT-4o) — secondary driver. Useful for image analysis, multimodal tasks, and as a cross-check on reasoning I want an independent take on.
• Perplexity AI — AI-powered search for research and citation-backed answers. Faster than manually cross-referencing Google for technical questions.
• Gemini 1.5/2.0 Pro — long-context document analysis. Useful when I need to ingest large technical specs, RFPs, or policy documents in one shot.

AI coding tools

• Claude Code — daily coding workflow (see above). The agent loop and MCP integration make it the most capable coding tool I’ve used.
• Cursor — AI-first code editor built on VS Code. Good for greenfield projects and large cross-file refactors. I switch between Cursor and VS Code + Claude Code depending on task complexity.
• GitHub Copilot — available in VS Code; useful for repetitive code completion but I prefer Claude Code for reasoning-heavy tasks.

Orchestration & multi-agent

• LangGraph — stateful multi-agent graphs with cycles, interrupts, and human-in-the-loop checkpoints. My default for production AI workflows that need durability and control.
• LangChain — orchestration scaffolding for simpler pipelines where LangGraph is overkill.
• CrewAI — role-based multi-agent framework. Good for sequential agent pipelines and team-simulation patterns.
• Model Context Protocol (MCP) — Anthropic’s open protocol for connecting AI systems to external tools and data sources. I run several MCP servers (Google Ads, Google Calendar, Gmail, Slack, browser automation, Canva). It’s how I’ve started wiring AI into actual operational workflows rather than isolated chat prompts — the difference between a tool and an assistant.
• AutoGen (Microsoft) — multi-agent conversation library; used selectively for prototyping conversational agent patterns.

RAG & vector infrastructure

• Weaviate — vector database of choice. Runs cleanly on-premise which matters in government-sector work. Clean Python client, good hybrid search support.
• LlamaIndex — data framework for LLMs; better than LangChain for document-heavy RAG pipelines that need careful chunking, metadata filtering, and retrieval evaluation.
• Ollama — local LLM runner. I run Llama 3.1, Phi-3, and Qwen2 for prototyping where data residency constraints prevent sending data to external APIs.
• Pinecone / Qdrant — managed and self-hosted vector databases respectively, for projects needing purpose-built infrastructure.
• Hugging Face — model discovery, embeddings, and fine-tuning experiments. The go-to index for model evaluation.

Observability & evaluation

• Langfuse — open-source LLM observability. Traces, evals, prompt versioning. The tool I wish I had two years ago when debugging production AI pipelines without visibility.
• OpenRouter — unified API across LLM providers. Useful for comparing models on identical prompts in production without re-wiring integrations every time a new model ships.

Workflow automation with AI

• n8n — self-hostable workflow automation. Increasingly use it to wire AI steps (classification, extraction, summarisation) into operational processes without building bespoke backends from scratch.

Project management & delivery

• Jira - most engagements use it. I’m fluent.
• Microsoft Office (Word, Excel, PowerPoint) - the lingua franca of executive communication. Don’t fight it.
• Notion - personal knowledge management and program documentation.
• Miro / Lucidchart - architecture diagrams and stakeholder mapping.
• Loom - async walkthroughs are underrated.

Writing & content

• Obsidian - local-first markdown notes. Where ideas mature before they reach the site.
• VS Code - for writing this site’s MDX.
• Grammarly - light copy-edit for English writing.

Web development for this site

• Astro - static site generator. Zero JavaScript by default, content collections, MDX. The right tool for a writer-first personal site.
• Tailwind CSS - utility-first styling.
• Pagefind - build-time search with no backend.
• Self-hosted Google Fonts via @fontsource - Fraunces, Inter, JetBrains Mono.
• GitHub for version control. Hosted on shared hosting via FTP - proof that you don’t need fancy infrastructure to run a fast personal site.

Browsers

• Arc for daily browsing
• Firefox for testing and privacy-sensitive work
• Edge for compatibility checks and Microsoft 365 work
• Chrome for DevTools when nothing else will do

Productivity

• Apple Notes / Microsoft OneNote - quick capture
• Calendar: Outlook for work, Google Calendar for personal
• Email: Outlook (work), Gmail (personal)
• VPN: [TBD]

What I don’t use

A list almost as important as what I do use:

• I don’t use AI image generators for production content. The site has no AI-generated imagery.
• I don’t use heavy IDEs (IntelliJ, VS) when VS Code can do the job.
• I don’t use no-code site builders for personal work - too much lock-in, too little control.
• I don’t use Twitter/X for thinking out loud anymore. Long-form writing is a better signal.

This list will go out of date. I’ll refresh it every few months. If you spot something interesting and want to compare notes - particularly on the AI tooling - I’d love to hear what’s working for you.