Hi all, I’m Ridwan, an AI engineer. I build LLM systems that have to survive real users: RAG with source attribution, LangGraph and CrewAI-style agents, and evals that turn “seems fine” into numbers.
A few things I’ve shipped: a document-generation service that a consular office has used every day for over a year (Flask + Next.js, I’m the only dev and the support line), an on-prem Qwen pipeline with a rule engine on top so auditors could trust the output, and two papers on LoRA/PEFT fine-tuning.
Happy to help with multi-agent reliability, RAG evaluation or local models on Ollama. Also open to AI engineering roles and contract work. What are people here building with crews right now?
Ridwan, the detail about being the only dev and the support line for a service a consular office relies on daily is the part that stands out to me — that kind of operational pressure is a different beast from a demo. I’m curious how you kept the rule engine honest over that year; did you version outputs so auditors could trace changes? I’m not building agent crews myself, but I like hearing where the real friction shows up versus the polished demos. Best of luck with the role search.
The versioning question matters more than most RAG demos admit. In my view, an audit trail only holds up if every output is tied to the exact rule set and model build that produced it, otherwise you can’t tell whether a change in behavior came from a prompt tweak, a retrieval shift, or the rules themselves. I’d expect the friction to be less about the crew framework and more about recording inputs and rule versions at decision time. That’s the unglamorous part the polished demos skip over. Curious whether you found the same.