Building AI Agents That Actually Work — From Scratch, In Production
I write about building real AI systems, the mechanism, not the marketing. No “five lines of code and you have an agent.” I build things by hand first, break them, fix them, and only then bring in a framework, so every abstraction gets evaluated against something I actually understood first.
Around three decades writing software, including running a ~30-engineer delivery organization inside a regulated fintech company, is the lens everything here gets filtered through: does this hold up in production, under compliance, at scale, not just in a demo.
What you'll find here
Builds, not tutorials. Full walkthroughs of systems built layer by layer, tool calling, agent loops, memory, retrieval, live web access, session persistence, with the actual bugs, actual GPU debugging sessions, and actual dead ends included. If something broke, it's in the post.
Framework comparisons that mean something. I don't review frameworks by reading their docs. I build the same system twice, once by hand, once with the framework, and show exactly what got replaced and what didn't. That's the only way to tell if an abstraction is saving you time or just hiding a decision you didn't know you were making.
On-prem and fintech-aware. A lot of AI content assumes you can call an external API with your data. I work in an environment where that's not always true, so a recurring thread here is what's possible entirely on your own infrastructure: local models, local compute, zero data leaving the machine.
Latest: I Built an AI Agent From Scratch, Then Rebuilt It in LangGraph
A full, hand-built AI agent, tool calling, the agent loop, memory, RAG, live web access, running entirely on a local GPU with Ollama and Qwen3. Then rebuilt piece by piece in LangGraph to show exactly what a framework automates versus what it doesn't touch.
The code for both versions is open on GitHub: github.com/truepythoneer/local-ai-agent
Follow the build
New posts cover real systems as they're built, not roadmaps or predictions. If that's useful to you: