← The Lab
Exploration history
The full list — what I am working on now, and everything already worked through. Finished does not mean mastered; it means I got what I came for.
Passed, July 2026 Claude Certified Architect exam
Working through Karpathy’s Neural Networks: Zero to Hero — doing the exercises, not watching them
Building Strategies for Ed‑Fi API tool use by LLMs — still unsure whether MCP is the right shape or overkill
Open question Hallucination detection: current strategies, and whether LLM‑as‑judge holds up in production
Open question Output validation and evaluation for anything I would put in front of a customer
Learning to do Read benchmarks and leaderboards critically enough to know when a number means nothing
Done Anthropic Academy: Claude Code in Action
Done Anthropic Academy: Building with the Claude API
Done Anthropic Academy: Introduction to agent skills
Done Anthropic Academy: Introduction to Model Context Protocol
Done Build a Modern Computer from First Principles: From Nand to Tetris (Project-Centered Course), Hebrew University of Jerusalem
Done How GPT, Claude, and Gemini are actually trained and served – Reiner Pope, Dwarkesh Patel podcast
Done Adam Marblestone – AI is missing something fundamental about the brain, Dwarkesh Patel podcast
Done Masked Diffusion Models — Stanford CME295 Transformers & LLMs, lecture 9
Done MIT OpenCourseWare 15.773 Hands-On Deep Learning
Done Hallucination detection: semantic entropy
Done Hallucination detection: Bayesian estimation of semantic entropy
Done LLM tool use mechanics
Done Token efficiency with tool calling
Done LLM-hosted tool calling costs and implications
Done Forcing JSON or other schema responses by giving the LLM a ‘fake’ tool
Done Security considerations with LLM tool use
Done MCP server primitives: tools, resources, prompts
Done Nuances between function/tool calling and packaged LLM features (agents, memory, etc.)
Done Tokenization, byte pair encoding
Done Prompt caching
Done Token pricing: input, output, cached
Done Token count and cost differences by format (JSON, YAML, code, prose)
Done Designing LLM conversations for efficient cost
Done Developing and testing intuition on model selection (Claude Haiku vs Sonnet vs Opus)
Done LLM batch processing
Done Anthropic’s contextual retrieval offering
Done RAG pipeline and RAG types (embeddings, prompt RAG / table-of-contents pattern, ELITE, full-text search, graph RAG), vector databases, semantic search
Done Pipeline: query → decompose → run each → rerank summed results
Done Generating embeddings via Azure AI tools
Done Embeddings models, tradeoffs, and how to choose
Done Semantic math with embeddings