How lemonflow built real-time agents for EV charging using Gemma and TPUs
lemonflow deployed self-hosted Gemma 4 on Google Cloud TPU v6e in the EU to run real-time AI agents for EV charging support, handling over 99.9% of live conversation turns and resolving more than 80% of issues without human handover.
lemonflow builds AI agents for EV charging networks, offering 24/7 driver support in over 90 languages and resolving a driver query every 90 seconds. The company reports that between 15% and 25% of EV charging attempts in Europe fail, often leaving drivers stranded or waiting in call queues. To address this, lemonflow developed conversational agents that can restart chargers remotely, restart sessions, and provide real-time updates in the driver’s language within seconds of contact.
To meet ultra-low-latency requirements for voice support, lemonflow migrated from managed APIs to self-hosting the open Gemma 4 model on Google Cloud TPU v6e hardware in the EU. This shift provided direct control over the model hosting environment, enabling deterministic security controls, prompt injection safeguards, and standardized backend protocols. The move also allowed the company to implement Agent Coach, a background engine that uses live call data to improve future interactions and automate regression testing across thousands of scenarios.
lemonflow’s architecture prioritizes live voice calls, followed by web chats, with automated fallback routing to self-hosted GPUs and Gemini Enterprise endpoints during system updates. Custom optimizations to the open-source vLLM serving engine and memory tiering between TPU and host memory enable a single TPU slice to maintain many operator agents ready for instant responses. The system achieves a typical response time of around 350 milliseconds while screening inputs for malicious prompts to meet European critical-infrastructure requirements.
The platform operates two types of AI agents: real-time support agents powered by Gemma on Cloud TPUs for minimal latency, and background monitoring agents running on Gemini to analyze conversations and charger logs. A translation layer standardizes varied CPMS interfaces across charge point operators, e-mobility service providers, and charger manufacturers, ensuring consistent performance. Drivers receive immediate assistance, including remote charger restarts or session corrections, with the system designed to keep interactions invisible to the user while maintaining high reliability.