There are plenty of podcasts about the energy transition and AI. Not many go deep on the engineering layer where the two actually meet; the architectures, the trade-offs, and the lessons that only come from shipping real systems in a domain where data is messy.
That’s the gap we’re here to fill. AI x Climate Tech is a podcast series from Enjins for CTOs, Heads of Data, and senior engineers in climate and energy. Each episode is a 30–45 minute conversation focused on concrete system design.
Episode 2: Optimizing Energy Behind the Meter
For episode two, we go deep into the energy transition. Arjan van Rooijen, CPTO at Iwell, joins us to talk through what it actually takes to turn raw hardware data into smarter energy decisions, at scale, across markets, behind the meter.
We cover:
- Why the energy transition can only be solved locally and what that means for how you build your data and optimization stack
- How iwell manages second-by-second sensor data across complex multi-asset sites, and what still breaks
- The layered approach to optimization: from powering daily operations to cost reduction to active energy trading
- Where generative AI is already making a difference. From AI-assisted customer support to on-site decision-making agents
- Why Arjan believes every CTO should also think like a CPO, especially now
What’s next
We have strong conversations lined up, covering everything from productionizing agentic AI at scale, to real-time energy data handling, to how climate tech investors evaluate AI value propositions.
If you work in climate or energy and want to talk architecture, we’d love to hear from you.
TL;DR Recap of Episode 2
- The grid can’t keep up. Net congestion is now a hard reality — in the Netherlands, entire provinces have stopped issuing new connections. The answer isn’t a bigger central grid; it’s optimizing behind the meter, where the businesses, consumers, and the bulk of the assets actually sit.
- Good optimization starts with good data, and that’s harder than it sounds. Sites are rarely a single grid connection — they’re multiple transformers, different constraints, and sensor readings interrupted by communication dropouts. Understanding where every data point comes from, and how it’s measured, is foundational before any ML can add value.
- iwell runs local models on-site, monitoring on a second-by-second interval, while syncing with the cloud at 1–5 minute intervals to forecast and optimize ~36 hours ahead. If the connection drops, the site keeps running and optimizing locally — never exceeding its constraints.
- Optimization happens in layers. First, powering daily operations so customers can run and grow their business within their constraints. Then, doing it cost-effectively. Then, where capacity allows, trading the surplus. iwell deliberately stays trader-agnostic, opening APIs rather than locking customers in.
- Constraints differ wildly by market. Active vs. reactive power in Belgium, half-hourly vs. quarter-hourly pricing intervals, different grid codes and energy markets across the UK, Germany, Austria, and beyond. The pipeline structure stays similar; the constraints are what change.
- Generative AI is already in production — on two fronts. Internally, agents pull from customer data, Notion documentation, and Slack to give support teams optimal answers on the spot, enabling scale without doubling headcount. And increasingly, AI supports on-site optimization decisions, always underneath a hard deterministic layer that guarantees safety and compliance.
- iwell deliberately chose EMS hardware with more memory and CPU — not the cheapest embedded option — so it can eventually run local agent models directly on customer sites. Small models, no GPUs, but enough to make real trade-offs on the edge.
This episode is a deep dive into what it actually takes to optimize energy at that decentralized edge: the data, the machine learning, and the emerging role of agentic AI.
From EV charging to energy management
Arjan’s path into energy started about ten years ago. After building data platforms in the internet wave and selling a company to translation-technology leader SDL, he moved into EV charging. First as CTO at EV Box, later at Pod Point, building charging infrastructure at scale across Europe. That experience taught him a core lesson: it’s rarely about charging vehicles. It’s about charging them at the right moment, and eventually using them as a way to manage energy locally.
At Iwell, he now leads both product and technology, a combination he argues is essential. Good technology development, in his view, always starts from the customer perspective, and you do that best when product and engineering sit together.
The hardware/data game
Everything starts with data quality, and that’s harder than it looks. People assume sensor technology is precise and reliable. In reality, there are discrepancies, communication interruptions, and noisy readings everywhere. A single site is rarely a single grid connection. It’s multiple transformers, different constraints, different needs. You have to understand where every data point originates and how it’s measured before you can trust it.
Iwell’s customers are typically larger sites: transport companies charging fleets of trucks, big apartment buildings, construction sites. Many have invested heavily in solar, but net congestion means they often can’t expand their grid connection. So they have to do more with what they have. Iwell’s platform manages all the assets on-site (batteries, charging, PV, peak shaving), understanding the local constraints that, combined with the data, are the key to optimization.
The architecture is hybrid by design. Local models run on-site at second-level monitoring, handling constraints and optimization even if the cloud connection drops. The cloud layer forecasts and optimizes roughly 36 hours ahead, pulling in external data such as weather for PV forecasting and market tariffs across multiple markets, then syncs that intelligence back down to the local model.
What iwell optimizes for
The hierarchy is clear. The core is what Iwell calls “powering daily operations”: making sure customers can run and grow their business without being limited by their constraints. On top of that comes cost-effectiveness. And where there’s spare capacity, trading.
These three elements have to come together against a backdrop of rising complexity. Vehicles are becoming bidirectional, capable of delivering energy back to the grid, which sharply increases the constraints: how long will the vehicle be connected, can it still be recharged in time, when does it need to leave. Regulation is starting to catch up to support this kind of local flexibility, though Arjan notes the tax system still often works against it.
The structure of the technical pipeline stays roughly consistent across markets. The constraints don’t. In Belgium you optimize for reactive power as well as active power. The UK uses half-hourly pricing while much of Europe has moved to quarter-hourly. Grid codes, energy markets, and incentive programs differ everywhere. A model has to make the right decision locally, with the right mix of cloud data, local sensor data, and external context about the customer’s business.
Where agentic AI comes in
Most of what’s described above is machine learning and forecasting. But generative AI is now entering the picture on two fronts, and both are already in production.
The first is the workforce. The energy domain is dense with technical documentation, asset specs, and regulation. Iwell runs agents across customer data, Notion documentation, and Slack communication to give customer support teams the most accurate answers on the spot. The result: the ability to scale without doubling the support team, with humans still in the loop reviewing the information before it reaches the customer. It’s the safer, higher-confidence bet.
The second is more ambitious: applying AI to the optimization game itself. Here Iwell uses a strictly layered approach. At the lowest level sits the “energy orchestrator”, a hard deterministic check that guarantees every market and safety constraint is met. On top of that, AI in combination with machine learning is increasingly used to find new constraints and optimization opportunities, and to make the optimal trade-offs. The protective layer is never bypassed.
Looking ahead, Iwell deliberately specced its EMS hardware with extra memory and CPU, not the cheapest embedded option, so it can eventually run local agent models directly on customer sites. Not large language models, no GPUs, but enough to make real local trade-offs at the edge.
