Writing
LlamaIndex: Giving Agents Your Documents
The five-stage RAG pipeline: loading, indexing, storing, querying, evaluating. LlamaParse, chunking trade-offs, query engines as agent tools, and picking your framework.
LangGraph: Thinking in Graphs
Why explicit control flow beats free-roaming agents. State, nodes, and edges explained. Building the email-sorting butler graph with conditional routing, step by step.
LangGraph in Production: Agents That Survive
The ReAct loop as a two-node graph. Checkpointing that survives crashes. Human-in-the-loop interrupts, retries, fallbacks, and full observability with Langfuse.
The GAIA Capstone: Building an Agent That Passes the Final Exam
The Agents Course capstone: 20 GAIA level-1 questions, exact-match scoring, a 30% certificate bar. Architecture, the scoring API, answer formatting, and the pitfalls that eat scores.
The Future of Agents: Fine-Tuning for Function Calling, MCP, and What Compounds
LoRA fine-tuning for function calling with special tokens, PEFT and TRL, when fine-tuning beats prompting, multi-agent teams, the Model Context Protocol, and the field's open problems.
Agents in Games: What a Pokémon Battle Teaches You About Production AI
Games are the best agent laboratory ever built. Building an LLM Pokémon battle agent with smolagents, battle tools, an agent-vs-agent leaderboard, and the production lessons hiding inside.
Agentic RAG: Retrieval That Thinks
Why one-shot RAG pipelines fail, and what changes when retrieval becomes a tool the agent wields. The gala example, tool descriptions as routing logic, and hybrid search.
Agent Evaluation and Observability: The End of Vibes-Based Development
Why agents fail silently and how to see inside them. OpenTelemetry instrumentation, Langfuse traces, offline vs online evaluation, LLM-as-judge, and the GAIA benchmark.
Same app. Very different run.
Build the same app five times — once under each framework — and watch the cost, the time, and the codebase drift apart. A head-to-head of Claude Code native, Superpowers, OpenSpec, Spec Kit, and GSD, with the methodology honestly disclosed.
The Loop Engineering Commandments
Prompting gives answers. Loops ship work. Ten commandments from building Lexomat — a full-stack legal AI — in production.
When agents dream: out-of-band memory and the evolution of agentic memory
How agentic memory evolved — from CLAUDE.md to memory-as-a-filesystem, then from in-band learning to an out-of-band consolidation pass Anthropic calls 'dreaming' — with a production engineer's read on which parts are real leverage and which are classic database primitives in a trench coat.
Same talent. Different system.
Two engineers, the same talent. One is buried in ops and fixed deadlines with no room to start anything new; the other owns the project and just builds. The difference isn't talent — it's the system, and the two ceilings it quietly sets: how many experiments happen, and how long builders stay motivated.
What 'in production' really asks of an LLM system
A production LLM or agent system succeeds only when quality, latency, cost, feedback and fallback hold at once — and aligned with the product. That alignment is also where the big, safe wins hide: sometimes a single model swap is far cheaper for the same quality, if you can prove the quality didn't move.
Are you AI-transitioning your company? Awareness & availability ≠ adoption!
Your people know the tools exist and the licences are paid for — but the AI transition stalls when the reward system and planning don't change with it. Delivery planning shouldn't compete with exploring, using and demonstrating AI.
Smolagents in Production: Tools, Models, and Deployment Patterns
Model selection trade-offs (local vs API vs premium). Building multi-tool agents. FastAPI async patterns. Deploying Smolagents to production. Real-world considerations.
Building Code-First Agents with Smolagents
Smolagents philosophy. ToolCallingAgent vs CodeAgent. Building your first weather agent in 50 lines of Python. Complete code walthrough.
How LLMs Actually Work: Tokenization, Context, and Function Calling
What language models are really doing under the hood. Why tokens matter. How function calling solved the hallucination problem. The fundamentals that make agents possible.
Agents 101: What They Are and Why They Matter Now
The difference between a chatbot and an agent. Why function calling changed everything. How to understand the agent landscape in 2025.