The archive
Notes, experiments, and write-ups from the AhsanLab workbench.
A demo RAG pipeline is forty lines of code. A RAG system that gives correct, grounded answers to real users at real scale is a different animal entirely. Chunking, hybrid retrieval, reranking, evaluation, caching, and the guardrails that stop it from confidently lying.
Generative AI, an AI agent, and agentic AI are three different things — and most teams ship the wrong one because they never drew the line. A plain-language map of what each actually is, how they stack, and how to pick the right altitude for the job. With diagrams, no hype.
Adding a second WebSocket node is easy. Keeping Redis from melting at four nodes is the part nobody benchmarks. The fan-out amplification trap, the one-line subscribe mistake that costs you 30× the message volume, and real numbers from 1 to 8 nodes.
I built my production game on Node.js, then rebuilt the same broadcast benchmark in Django Channels and Phoenix to settle the argument honestly. Three concurrency models, three sets of real numbers, and a decision guide that does not pretend one framework wins everything.
The memory graph climbed all night while the connection count sat flat at 500. The culprit was not a code leak — it was zombie sockets the garbage collector could never reach. A war story, a reproduction, and the five-line fix.
Connection lifecycle, authentication, room design, backpressure, and Redis pub/sub — everything it actually takes to run a WebSocket backend under real user load. Drawn from Matrix Bingo, a live multiplayer game at scale.
