Published Paper Details:

Serverless computing (Function-as-a-Service, FaaS) is severely limited by the COLD start problem, a latency penalty incurred during container initialization. Current mitigation approaches, relying primarily on static keep-alive policies or single-source time-series prediction, fail to balance performance consistency with resource cost, often leading to a 40-60% waste in resource provisioning. We introduce COLDSTART (Cost- and Latency-Optimized Serverless Orchestration via Multi-Modal Prediction and Confidence-Based Warming), a novel, intelligent orchestration framework designed to minimize COLD start occurrences and resource overhead simultaneously. Our core contribution is a Multi-Modal Prediction (MMP) architecture that fuses six heterogeneous data streams--spanning user behavior, API metrics, and system telemetry--into a unified feature space. This space is analyzed by a tri-model Ensemble Learning Engine (XGBoost, LSTM, and Transformer) to generate highly accurate invocation forecasts. Crucially, we propose a Confidence-Based Warming (CBW) mechanism that gates resource allocation based on a calculated prediction confidence score (C), thereby minimizing expensive false positives. Through extensive experimentation on simulated production workloads, COLDSTART achieved a 68% reduction in COLD starts and improved 95th percentile latency to under 200 ms, concurrently delivering a 45% reduction in operational waste compared to single-source ML baselines. This work demonstrates a robust, context-aware solution essential for sustainable and high-performance serverless deployments. Index Terms--Serverless Computing, Function-as-a-Service (FaaS), COLD Start Mitigation, Multi-Modal Prediction, Ensemble Learning, Cost Optimization, Proactive Warming.