The objective of the proposed research is to enable reliable deployment of a range of deep learning models on analog crossbars leveraging post-manufacture testing and tuning. Memristive crossbar arrays (MCAs) have emerged as a promising technology for reducing inference energy of deep neural networks (DNNs). However, their widespread adoption is hindered by hard defects as well as device-level conductance variation within the crossbar. As a result, when an artificial intelligence (AI) model is deployed using MCAs, its performance (classification accuracy for an image classification task) degrades. This research aims to reliably deploy a range of AI models such as convolutional neural networks (CNNs), spiking neural networks (SNNs), generative adversarial networks (GANs) and small language models (SLMs) on analog crossbars using fast and accurate post-manufacture testing. To assess the extent of performance degradation of each manufactured chip and discard the chips with unacceptable degradation, adaptive functional test frameworks are designed for SNNs, CNNs, GANs and SLMs. Further, for privacy-critical applications, a dataset-agnostic testing framework using synthetic data is developed. Discarding MCA-based chips due to hardware non-idealities reduces manufacturing yield. This research addresses yield recovery using hardware-friendly lightweight post- manufacture tuning. A small set of tuning knobs is optimized for every chip to recover performance loss. Experiments show that the proposed framework achieves less than 1% test escape and improves yield by 34%.