I built DataSentinel AI, a Python‑based monitoring stack that watches the data feeding my machine‑learning models. It runs a series of quality, schema and drift checks, then alerts and launches an investigation agent that composes an incident report with a free local LLM via Ollama.
The repository language is Python.
The project started in 2026‑10 and was last updated in 2026‑10.
I use the free local model qwen2.5:3b via Ollama as the LLM backend.
A feature is considered drifted when PSI ≥ 0.2, or when KS p < 0.01 together with PSI ≥ 0.1.
Evaluation on 2026‑10‑09 showed detection accuracy of 1.00 for both rule‑based and live‑LLM modes.
False‑positive rate was 0.00 for both modes.
Mean alert‑to‑report latency was under 0.01 s for rule‑based and 40.7 s for live LLM.