Nepal’s complex mountainous topography renders its river basins highly susceptible to devastating flash floods, yet existing monitoring infrastructure is often cost-prohibitive and technically limited in remote hilly regions. This paper presents a terrain-adaptive, real-time Flood Early Warning System (FEWS) utilizing a robust three-tier IoT architecture. The system employs ESP32 microcontrollers interfaced with JSN-SR04T waterproof ultrasonic sensors for non-contact water level detection. A key innovation of this research is the integration of a dedicated hill-top repeater station to overcome the Line-of-Sight (LoS) limitations of LoRa communication in non-linear, mountainous terrains. To ensure signal integrity, a 10-second median filtering algorithm was implemented, effectively eliminating false triggers from surface ripples or debris. Furthermore, a delta-based transmission strategy was utilized, reducing redundant MySQL database load by 98.2%. Experimental validation conducted across a multi-block campus environment confirms that the system achieves a 95% packet success rate in non-line-of-sight conditions and triggers multi-channel emergency alerts (SMS, Email, and localized Siren) with minimal latency. The proposed solution offers a low-cost, scalable, and power-efficient alternative for disaster risk reduction in the flood-prone watersheds of Nepal.