I write The AI Engineer, twice-weekly breakdowns of agents, inference and GPUs, read by 35K+ engineers at NVIDIA, Anthropic and OpenAI. I'm also writing AI Inference Engineering for Manning.
I spent 8 years as an ML engineer, most of them building Prediktiva, the first data science agency in Latin America, which I founded and which was acquired in 2024. Before ML, I started out in investment banking.
Contributor to:
Start here: Latest Issues · The Book · Work With Me · Open Source
- What is Agentic Search?
- How to Build an Agent Eval You Can Trust
- Mac Studio vs DGX Spark vs RTX 5090
- What is a Transformer?
- Continuous Batching vs Speculative Decoding
My book with Manning on serving LLMs in production: how a request becomes tokens and what every layer costs, from GPUs and kernels to batching, KV cache and speculative decoding. Companion code: ai-inference-engineering.
1M+ AI engineers read my content. NVIDIA has sponsored it since 2023, and 100+ AI/ML companies have followed. Yours could be next.
 AWS |
 Microsoft |
 |
 NVIDIA |
 Salesforce |
 MongoDB |
 Redis |
 LangChain |
 LlamaIndex |
 CodeRabbit |
 Fireworks AI |
 Sonar |
Most repos are the companion to an issue of The AI Engineer: one failure from production, made small enough to run in a minute.
| Repository | What it shows you |
|---|---|
| fraud-foundation-vs-boosting | Whether a tabular foundation model beats tuned XGBoost on fraud, and what it costs to serve. Runs on a free Colab T4. |
| boxfit | Whether your model fits an RTX 5090, a Mac Studio or a DGX Spark, and how fast it runs. |
| spec-window | The batch size where speculative decoding stops paying on your vLLM server, from one /metrics scrape. |
| gpt2-by-hand | GPT-2 small in plain numpy, block by block, matching Hugging Face to within 0.0007. |
| self-host-boundaries | A self-hosted LLM stack broken on purpose, one failure per boundary, to test your pre-flight checks. |
| agent-harness | The 5 layers of an agent harness, each with the failure it prevents. Runs with no API key. |
| agent-crash-recovery | The same agent on LangGraph, Inngest, DBOS and Temporal. Kill it mid-run and see which steps you pay for twice. |
Smaller tools: cuda-stack-check (the layer breaking your PyTorch CUDA stack) · kernel-trace-check (the one lever your step time points at) · chunk-position-check (the prompt slot your model reads worst) · transformer-xray (context or training: where an answer came from) · rl-post-training (attack any grader before it decides anything) · red-green-check (tells a proven bug fix from a flagged one) · compaction-audit (what survives when Claude Code compacts your conversation) · ai-engineer-reading-list (a daily.dev plugin that writes your weekly AI engineering brief)
Everything here is free to read and run. Three ways to keep it that way:
- Star the repos you use. It's how other engineers find them.
- Share a tool or an issue with your team.
- Sponsor the work through the Sponsor button on any of my repos, or partner with me on a piece.