Work

Closed-Loop LLM-Driven Self-Healing for Enterprise Data Pipelines

Year: 2026
Type: conference-paper
Abstract: Enterprise data pipelines fail in recurring ways: schemas drift, files arrive corrupted, distributions shift, resources run out, and upstream sources go down. Large Language Model (LLM) agents have shown that they can diagnose such incidents accurately, but published systems stop at advising a human because LLMs also exhibit systematic reasoning failures. We present a closed-loop architecture that lets an LLM controller observe, diagnose, propose, verify, and remember, but only commits a remediation that passes a three-stage check (static, replay, and a statistical shadow test based on a Kolmogorov–Smirnov bound). The action space is closed and typed, so the LLM cannot invent a new fix. We instantiate the architecture in SHIELD-Bench, an open reference harness, and run a preliminary 300-incident study across five configurations and twelve fault types in eight classes, using both a deterministic rule-based diagnoser (included as a reproducibility baseline) and a real LLM diagnoser. Key results: the LLM diagnoser reaches 0.778 overall accuracy (six of eight classes at 1.00); the verifier catches every unsafe remediation in the evaluated library (recall 1.00 under both diagnosers; precision 0.667 rule-based / 0.632 LLM); closed-loop verification gives up 8.3–11.7 percentage points of auto-resolution to obtain empirical predicate-bounded safety evidence over the implemented checks. The adversarial study sharpens this boundary: open-loop auto-deploys every harmful remediation, while the shadow check intercepts all of them before any data is written. Compositional failures remain unresolved but are correctly escalated. LLM cost is $0.0019 per diagnoser call. The harness, prompts, seeds, and results are released as supplementary material.
Language: English
Open Access status: hybrid
OpenAlex ID: W7214743239
FWCI:
Cites: 15
Cited by:
Related to: 10