Navigating AI-First Products: Responsibilities, Competencies, Obstacles, and Tactics for AI Product Managers
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Abstract Production artificial intelligence systems suffer from critical reliability gaps due to non-stationary data distributions causing concept drift, stochastic inference latency, and unbounded explainability overheads that violate service level agreements. Current MLOps tooling lacks statistically-grounded mechanisms for proactive drift management with quantifiable false positive constraints, resulting in mean time to recovery (MTTR) exceeding 34 hours and 22.7% SLA violation rates under load spikes. This paper develops and validates DriftGuard, an open-source engineering framework comprising three integrated components: (1) DriftQuant Engine implementing ensemble Population Stability Index and Kolmogorov-Smirnov detection with operator-derived thresholds (PSI > 0.25 sustained 72 hours) that reduce false positives to 4.20%; (2) StochBound Protocol employing quantile regression coefficients (β_velocity = 0.73, β_queue_accel = 0.41) for predictive resource allocation that maintains p99 latency within ±10% of targets during 3× traffic surges; and (3) XAI-Inject Pipeline enforcing strict 200ms latency budgets for SHAP explanations with 94.7% compliance. Rigorous validation across computer vision (CIFAR-10), NLP (IMDB), and predictive analytics (NYC Taxi) workloads demonstrates 30.4% MTTR reduction (34.2 → 23.8 hours, p = 0.0003, Cohen's d = 0.92), 42.3% SLA violation reduction, and 30.1% model decay suppression versus manual baselines. The Dockerized artifact enables immediate adoption for hardening production AI systems against non-stationary degradation while meeting IEEE fields of Computing and Processing, Signal Processing, and Robotics and Control Systems requirements for applications-oriented engineering contributions.
Fields of Study & Subjects
Software System Performance and Reliability
Data Stream Mining Techniques
Cloud Computing and Resource Management
Publication details
- Journal / source
- Research Square
- Publication date
- 26 March 2026
- Language
- EN
- Document type
- Preprint