Artificial intelligence is rapidly reshaping education by making learning experiences more personalized to individual needs. AI systems can engage in natural language interactions with learners, respond to their questions in real time and provide immediate feedback. This not only supports deeper understanding but also makes the learning process more interactive and meaningful. Many of today’s widely used AI tools are proprietary, cloud-hosted platforms that require paid subscriptions. These solutions often raise concerns about cost and data privacy. An alternative approach is to use open-source small language models (SLMs) that can be deployed locally or on edge devices. Such models offer an affordable option while ensuring that data remains secure. In this paper, we examine different strategies for building AI-driven learning assistants. Specifically, we compare open-source versus proprietary tools, small versus large language models, and code-first versus low-code/no-code development approaches. We also explore two common methods for expanding a model’s knowledge base: Retrieval-Augmented Generation (RAG) and fine-tuning. Each approach is evaluated in terms of cost, privacy, accessibility, and the technical expertise required for implementation. We present a prototype learning assistant built using a practical combination of open-source low-code/no-code development environments, RAG and locally deployed open-source SLMs. The assistant can provide relevant, context-aware responses and can be easily customized by non-technical educators and learners. By adopting cost-effective, privacy-conscious, and accessible technologies, it becomes possible to create AI solutions that genuinely serve the needs of diverse learning communities while reducing dependency on commercial AI providers.