A UNIFIED PIPELINE ARCHITECTURE FOR ENTERPRISE-READY LLMs: ADAPTIVE INGESTION, CONTINUAL LEARNING, SLO-AWARE RETRIEVAL, AND AUTOMATED FEEDBACK CONVERGENCERaza, Muhammad FaizanLarge Language ModelsLLM Pipeline ArchitectureRetrieval-Augmented GenerationContinual LearningCatastrophic ForgettingAdaptive IngestionSLO-Aware RetrievalFeedback OrchestrationError FingerprintingSelf-CorrectionLoRA AdaptersSparse Temporal Adapter RoutingKnowledge StalenessHallucination MitigationMAPE-K Control LoopStream ProcessingApache KafkaNeighborhood-Distilled Correction ExemplarsEnterprise AILLMOpsData AnalyticsMSLarge Language Models have rapidly transitioned from research prototypes to indispensable components of enterprise systems across finance, healthcare, emergency response, and real-time analytics. Yet the operational realities of production deployment expose four fundamental limitations that current architectures fail to address in combination: knowledge staleness caused by static pretraining cutoffs, catastrophic forgetting during parameter updates, hallucinated outputs that undermine factual reliability, and weak feedback mechanisms that allow systematic errors to persist without automated correction. Addressing these limitations in isolation, as the prior literature has almost exclusively done, is insufficient. Each failure mode interacts with and amplifies the others: stale knowledge makes retrieval less effective, imprecise retrieval makes hallucination more likely, and the absence of a closed-loop correction mechanism means that neither failure is caught before it reaches the user. This thesis takes the position that these four problems constitute a single engineering discipline and proposes a unified, six-layer pipeline architecture grounded in classical software design patterns to address all four simultaneously. The research is organized around four guiding questions. The first asks which real-time ingestion architecture, among Pure-Stream, Lambda, and CQRS patterns, best minimizes the end-to-end latency from data arrival to LLM-answerable freshness, leading to the Adaptive Ingestion Pattern Orchestrator (AIPO) and the Answerability Tracing Protocol (ATP), the first reproducible benchmark for measuring time-to-answerable freshness in LLM pipelines. The second asks how to update model parameters continuously without catastrophic forgetting under realistic compute and memory budgets, leading to STAR+FAR: Sparse Temporal Adapter Routing combined with Freshness-Aware Replay, a continual learning regimen that routes each query through a sparse weighted combination of per-day LoRA adapters while maintaining a hardness-novelty-recency-prioritized replay buffer of 128 examples. The third asks how retrieval-augmented generation can be optimized to meet strict service-level objectives simultaneously on accuracy, latency, and cost, leading to SAGE: a lightweight SLO-aware retrieval budget policy that learns, via oracle-guided imitation, to dispatch each query to its individually optimal retrieval depth k rather than applying a fixed budget to all queries. The fourth asks whether a closed-loop feedback controller can learn from oracle-demonstrated remediation choices to produce net-positive quality correction on real LLM outputs without human intervention, leading to CAFO: the Convergence-Aware Feedback Orchestrator, which combines error fingerprinting, neighborhood clustering, and a LightGBM selector trained on oracle decisions to issue targeted remediations that are validated by a sequential A/B test before being committed. The experimental campaign spans 1,620 pipeline variant runs across four LLM families, three benchmark datasets, and two hardware scales, accumulating approximately 400 A100 GPU hours. The results verify all four hypotheses. On RQ1, Pure-Stream achieves a P95 time-to-retrieval-ready latency of 117 milliseconds, 4.3 times lower than the CQRS baseline of 499 milliseconds, and AIPO improves burst resilience by 38 percent over any static architecture. On RQ2, STAR+FAR simultaneously achieves a 4.1 percentage point gain in factual freshness and a 3.2 percentage point gain in legacy retention relative to the best prior baseline, a Pareto improvement that no single-mechanism continual learning approach achieves, confirmed at both TinyLlama 1.1B on CPU and Qwen2.5-7B on GPU at 1.7 minutes per day update cost. On RQ3, SAGE achieves 95 percent SLO compliance compared to 30 percent for the best static retrieval configuration, a 36 percent reduction in P95 latency, and a 51 percent reduction in retrieval cost, while generalizing zero-shot across all four LLM families tested. On RQ4, CAFO achieves the only non-negative mean cumulative correction efficiency among five feedback regimens on the primary benchmark, and its trained selector transfers to a held-out model architecture with a cumulative correction efficiency of +0.329, the strongest single corrective result in the campaign, demonstrating that error fingerprint patterns encode failure-mode structure that is largely model-agnostic. Taken together, these contributions constitute the first unified, empirically validated framework for real-time enterprise LLM pipelines in which each architectural layer is optimized not merely in isolation but in explicit service of its neighbors, creating a virtuous cycle in which faster ingestion sharpens the continual learning signal, more accurate parametric knowledge reduces the retrieval burden, cost-efficient retrieval liberates headroom for denser feedback correction, and net-positive automated feedback closes the quality loop that all three preceding layers leave open. The work elevates LLMOps from ad-hoc engineering to a principled, reproducible discipline, delivering a strategic pipeline blueprint that maps business constraints directly to empirically validated architectural choices for practitioners building systems that must be fast, fresh, factual, and self-improving, all at the same time. Satish Mahadevan Srinivasan, Thesis Advisor/Co-AdvisorChengfei Wang, Committee MemberRaghu Sangwan, Program Head/ChairHajime Shimao, Committee MemberDusan Ramljak, Committee Memberopen_access2026-03-25T03:24:05Zhttps://etda.libraries.psu.edu/catalog/35191mfr5933