19 articleson building production systems — what worked, what didn't, and why.
RSS feedAgentic AI
Hardcoding fake data in an MCP server is fine for a demo. Making real, asynchronous REST API calls to external services is where the magic actually happens. Here's how to pair FastMCP with HTTPX.
AI Engineering
How self-attention lets a model decide which words matter to each other — the query/key/value mechanics, scaled dot-product attention, multi-head attention, and causal masking, with a from-scratch PyTorch implementation.
Agentic AI
Every one of these cost me real debugging time. Each has a one-line fix — here they are, with the code.
Agentic AI
All three get called 'AI coding assistants,' but they're built on genuinely different philosophies. Here's the real difference, in plain terms, with an actual terminal transcript.
Agentic AI
A weekend project, a bash tool with no restrictions, and the exact three mistakes that let it happen — plus the guardrails that would have stopped every one of them.
AI Engineering
Start from a plain sentence, tokenize it with NLTK, and build up through every major embedding type — one-hot, TF-IDF, Word2Vec, GloVe, and modern contextual embeddings — with working code for each.
Agentic AI
Almost nobody publishes real token math for multi-step agents. Here's a worked example — a 6-turn agent loop, priced out — plus a small calculator function you can reuse.
Agentic AI
Agents sound complicated until you see the actual request-response loop underneath. Here's the whole thing, stripped down to one tool and 20 lines of Python.
Agentic AI
One agent with a huge system prompt and thirty tools gets confused fast. Splitting it into focused agents that hand off work to each other usually works better — here's the pattern, with code.
AI Engineering
A practical, from-scratch walkthrough of PyTorch — tensors, autograd, building a model, training it, and where to go next — with links to the official docs and tutorials.
Agentic AI
RAG and agentic RAG get used interchangeably constantly, and they're not the same thing. The difference is one word: whether the model can decide to look again.
AI Engineering
Five terms that show up in almost every ML pipeline, explained simply and with code: word/sample reweighting, time scaling, data rescaling, smoothing, and bagging.
Agentic AI
Strip away the hype and every agentic system — Claude Code, MCP-connected agents, multi-step research bots — is running the same five-line loop underneath. Here it is, in full.
Agentic AI
MCP explainers are usually either a marketing slide or a 40-page spec. Here's the plain version — one analogy, one 15-line server — and why it matters more than it sounds like it should.
AI Engineering
A deep dive into word embeddings: from one-hot encoding limitations and TF-IDF matrices to Word2Vec skip-gram mechanisms, GloVe factorization, and contextual Transformer embeddings — with Python code for each.
Agentic AI
Memory is the first thing everyone asks about agents, and the first answer is always 'embeddings and a vector DB.' Here's the version that fits in a JSON file, and why you might not need more than that.
AI Engineering
Lessons from moving retrieval-augmented generation from a demo to something that survives real traffic and real data.
AI Engineering
Why traditional unit tests fall short for agentic systems, and what a useful evaluation harness actually looks like.
AI Engineering
A grounded look at what MCP actually solves, and where it fits in a real AI product's architecture.