Agentic Energy Reporting Platform, Kasra Zamanian

Setup

The Smart Energy Report was a GenAI-powered product built on the electrical-energy data from an IoT device developed by Huna, an early-stage startup that had recently joined Corporate Hangar. The device was a small, round unit that clamped onto a building's main electricity cable in a completely non-invasive way, and from that single point it captured every detail of consumption, even disaggregating the flow into sub-sections: individual rooms, specific machines, and so on.

The data was a goldmine for anomaly detection and for technical users. The problem was everyone else. For a non-technical building owner the readings were impossible to interpret, and turning them into prioritized, actionable insight took an expert data analyst with real field knowledge. That does not scale.

Why it was hard

The spark came from the CEO, an engineer himself. He had taken one installation's data, pasted it straight into ChatGPT, and gotten back a slick report of findings and anomalies. He was thrilled: the model had done, in one prompt, the analysis we had been planning to build with bespoke algorithms. He reached out to me, and since I was already deep in this data (working on anomaly detection and energy disaggregation, identifying which machines were running from their consumption patterns), I was the natural person to take it further.

I was excited and immediately suspicious. What people forget about LLMs is how desperate they are to please you, happy to hand over confident, fluent analysis with no factual basis whatsoever.

The first question: is it even idempotent?

What happened: Before celebrating, I tested the one thing that actually mattered for a report someone would act on: reproducibility. Same data, same prompt, run it again. The findings came back completely different. To no surprise of mine it was not idempotent, and an analysis that changes every time you run it is worthless. But one thing stood out and stuck with me: how effortlessly it had made genuinely complex analysis readable. That was the lamp going on.

The fork: Write the LLM off as an unreliable toy, or keep its one real gift, explanation, and engineer away its unreliability.

My call: Keep the magic, kill the hallucination. And the way to do that was to change what the model was allowed to touch.

What it led to: A completely different architecture, one where the LLM never sees the raw data at all.

Don't feed it the data. Feed it the tools.

What happened: Instead of dumping raw telemetry into a prompt, I gave the model the deterministic tools I was already building, my anomaly-detection and disaggregation algorithms, exposed as callable functions, plus context about each installation: was it a factory, a kitchen, a residential block, a school? The numbers only become findings in context. High consumption at 2am is perfectly normal for a 24-hour kitchen and a genuine anomaly for a factory that runs day shifts only.

The fork: Let the model analyze the data directly (fast, fluent, ungrounded), or constrain it to orchestrate trusted tools and reason only over their verified outputs (more engineering, fully grounded).

My call: Grounding over magic, every time. I orchestrated it with LangChain: the LLM became a reasoning layer on top of a deterministic tool layer, calling my algorithms and reasoning over their outputs, never inventing a number itself. The installation's context came from RAG: I retrieved its profile and the norms for its sector (what "normal" looks like for a hospital versus a school) so the model reasoned against real reference points instead of its own assumptions.

What it led to: Findings that finally had a factual basis. The explanations were as clear as the CEO's original ChatGPT demo, but this time they were true.

Impact

The pipeline ran on a schedule across more than 50 installations spanning very different sectors, from schools to factories to hospitals, generating a Smart Energy Report for each in the background. Because every installation shared the same instructions, schema, and sector knowledge, prompt caching kept cost and latency down as it scaled. Non-technical owners finally got the prioritized, plain-language insight that previously needed an expert analyst, at a scale no analyst could match.

Takeaways

Hand an LLM raw data and it becomes a confident liar. Ground it: give it deterministic tools and context, not the data itself.
Context is what turns a number into a finding. The same reading is routine for a kitchen and an alarm for a day-shift factory.
Structured output is not a nice-to-have. It is the backbone of an agentic system you can actually trust and reproduce.
Reliability is layered, not a single trick: retrieve real context (RAG), constrain the shape (structured output), and verify the claims (a fact-checker). Each catches what the others miss.