Agentic Workflow thats Self Improving for Repo Understanding

Most "AI dev tooling" demos stay confined in a box and may rely on distillation or other mechanism. This one is a closed-loop system that improves the exact model that writes its own documentation:
a small (3 B) local LLM gets tool access to a real repository's knowledge graph,
batch workflows generate, then fact-check and score, that repository's wiki,
the judged output becomes training data,
a fine-tuned release of the model is auto-deployed back into the serving stack,
every step is logged (MLflow, Jaeger, Prometheus) and gated by a human when it matters.
It deliberately does all of this on one 4 GB consumer GPU. No cloud, no API bill, no 70 B placeholder that could never run here — and the honest part is that the small-model constraints are why the architecture looks the way it does. Please refer resume or contact 0458793097 for further details