Notes, essays, and open-source AI projects by Moss GU.
An abstraction is a contract over a hidden implementation. Software stacks them for two reasons: nobody holds the whole system, and parts must change independently. Contracts came in two kinds, formal for machines and natural for humans, and an LLM is the first machine that consumes the natural kind in general. So intention is a new abstraction: you state the result, the model supplies the…
Clean and simple are two different questions, not two words for good code. They also work differently across a system's levels. Clean means the same thing at every level; simple only has meaning at one level at a time. And simplicity rests on a clean boundary. Clean itself has two layers: surface and structural. AI made surface clean almost free. It left you the rest: honest boundaries and…
Prompt engineering has a folk version — magic phrases, ever-longer instructions — and a real one. The real one is engineering. Reference concepts the model already shares instead of describing from scratch. Move down the six escalation levels only when a failing test forces you, and back up when a better model arrives. A prompt file is software, with the same forces and the same principles. And…
A demo proves an AI feature can work once. Production asks how often. You can't know that in advance, not from a benchmark and not from the marketing page. So try before you develop: set the bar as a number, make the try an eval against real cases, and pay for complexity only when the evidence forces it. Two real tries close the post — one that said yes, one that said no.
AI does development, not engineering — which is why 'entry-level' now means five years. The fix isn't hiring; it's designing above the code, where checking splits so juniors can verify again and the apprenticeship returns.
Design is Code (DisC) compiles PlantUML diagrams and decision tables into tests that pin the implementation — deterministic, reviewable AI code generation.
A knowledge base looks like a search box. Underneath it is a four-stage RAG pipeline where every stage fails silently. A comprehensive walk through the failure points — ingestion, retrieval, assembly, generation — plus the concerns that cut across all of them, how to evaluate the system, and where the field is going.
Vibe coding is for spikes; spec-driven development is for production. How specs, TDD, and agent orchestration make AI-generated code dependable.
Deploy a production-ready AI agent to AWS Lambda using OpenAI Agents SDK, FastAPI, and Terraform.
Eight AI orchestration patterns, from full control to full autonomy — with code examples and a framework for choosing how much AI complexity a problem needs.