Artificial intelligence no longer lives exclusively in the cloud. Increasingly, it needs to run directly on small, low-power devices such as sensors, instruments, and controllers close to where data is generated and acquired. This is the world of Edge AI, and FPGAs are emerging as an efficient option for achieving it, especially in applications requiring low latency, customized numerical precision, or resilience in demanding environments.

This talk shares real-world experiences deploying AI on FPGAs across very different fields, including radiation detection in nuclear instrumentation, crop monitoring in precision agriculture, and environmental monitoring. Behind these projects lies a common tool, hls4ml, which enables AI models to be translated into hardware-ready designs, making this technology accessible not only to hardware experts but also to researchers, engineers, and students from diverse backgrounds.

Building on lessons learned from both real-world deployments and educational activities, the talk examines recurring challenges encountered along the path from model design and compression to FPGA implementation. These challenges include fragmented toolchains, long design iterations, and the need for expertise spanning multiple domains. These experiences motivated the development of KalEdge, a hardware-aware platform that unifies dataset preparation, model design, compression-aware training, hardware-aware exploration, and FPGA deployment into a single end-to-end workflow based on hls4ml. The talk concludes with an overview of how KalEdge streamlines the machine-learning-to-bitstream process and helps lower the barrier to FPGA-based edge AI.