Welcome back to The Blueprint, a regular feature where we highlight how Google Cloud customers are tackling unique and common challenges across industries using the latest AI and cloud technologies. We hope to inspire others looking to innovate in their work.
As the electric vehicle revolution gains steam in Europe, every charging session is a moment of truth, for the driver and for the transition itself. At lemonflow, we build AI agents purpose-built for EV charging: support agents for drivers and monitoring agents for charging networks. Our support agents answer drivers around the clock in more than 90 languages, and today they help an EV driver every 90 seconds.
The Challenge
Before we founded lemonflow, we worked closely with leading mobility players across Europe and saw first-hand how often charging goes wrong: Between 15% and 25% of EV charging attempts in Europe end in failure. Across Europe’s rapidly expanding EV market, where battery-electric car sales grew 30% in 2025 and thousands of charge point operators run public chargers, drivers who encountered stalled charging points were often left stranded or stuck in slow call center queues.
We set out to change that hands-on: to bring the cutting edge of AI to EV charging, and to build the world’s best conversational AI for distributed hardware, an automated platform that resolves real-time charging issues via phone, chat, email, and WhatsApp. However, handling live voice support created several technical requirements:
- Strict latency thresholds: Frustrated drivers want immediate assistance. Our testing confirmed that serving latencies above 800 milliseconds risk conversational barge-ins, causing drivers to interrupt the agent and break the interaction flow.
- Managed API bottlenecks: Initially relying on managed LLM APIs, we faced latency deviations during peak hours (spiking from 1.0–1.2 seconds to over 1.7 seconds) and experienced HTTP 429 rate-limit errors when global usage peaked.
- Compute-intensive conversations: Every agent works from one comprehensive prompt, paired with dynamic tool calls that read and act on live charger data. This design creates intense bursts of compute demand during calls, making throughput on shared, managed capacity difficult to guarantee under peak loads.
- Highly fragmented ecosystem: We needed a standardized conversational engine capable of bridging around 30 distinct Charge Point Management Systems (CPMS), the backends operators use to run their chargers, each with its own APIs, data models and quirks.
- Hundreds of tailored agents: No two operators run support the same way. Each has its own tariffs, payment methods, escalation paths and hours for human backup, and most run several agents, for example one per country, brand or hotline, with behavior that can change by time of day. Each agent’s prompt is unique and must be served at full speed and improved without breaking any of the others.
To deliver ultra-low-latency voice interactions and guaranteed serving capacity, we needed direct control over our model hosting environment.
The Solution
We wanted every driver to get best-in-class support that feels effortless, almost magical: no queue, no menu, just a short conversation that ends with the car charging. Picture a driver on a long cross-border trip, pulling into a motorway service station at 11 p.m. She plugs in, but the charger won’t start: the previous session never closed properly. She calls the number on the charger and is answered in her own language. Within seconds, our agent spots the stuck session in the charger’s logs, restarts the charger remotely and stays on the line until power is flowing.
To get there, we had to solve our latency bottlenecks and secure dedicated compute capacity. We migrated from managed APIs to self-hosting the open Gemma 4 model on Google Cloud TPU v6e hardware in the EU.
Like many startups using Gemma, the compatibility of these open models with Gemini allowed us to transition our existing prompt infrastructure with minimal modification.
Moving to self-hosted TPUs unlocked the more complete architectural control our engineering team required to establish a deterministic core layer that handles EV-specific guardrails, guards against prompt injections, and standardizes previously fragmented backend protocols into our internal format before ever reaching the LLM. Control had a second payoff: because the model and serving stack change only when we change them, every test result stays meaningful. That stability laid the foundation for Agent Coach, our background coaching engine for AI agents, described below.
We then developed an intelligent routing gateway to prioritize workloads according to operational urgency, serving live voice calls first, followed by web chats, and deferring automated re-simulations to low-traffic periods.
Custom optimizations to the open-source vLLM serving engine and memory tiering between the TPU and host memory let a single TPU slice keep many more operator agents ready for instant answers, without recalculating their context.
We also implemented automated fallback routing to self-hosted GPUs and Gemini Enterprise endpoints. Thanks to TPU reliability, Gemma answers more than 99.9% of all live conversation turns, and these fallbacks serve primarily as deployment buffers during system updates. Primary or fallback, every request is served from secure EU compute.
