LaunchPrefaceVision

Introducing Grid SensAI

Making UK grid data accessible through natural language. A look at why we built this and how it works under the hood.

Sai Amith Chillamcherla11 January 20255 min read

Preface

As a grid engineer, my day-to-day job involves two main tasks. First, I manage grid connection applications for solar and storage projects, which means coordinating with DNOs and TSOs across Ireland and the UK. Second, I build Python-based tools to automate data analysis and power system studies.

Navigating multiple DNOs' data portals to find clean, standardized data has always been a struggle. I often wished for a tool that could just retrieve the clean data for me. That would allow me to focus my energy on the actual scripting, reporting, and analysis.

I have seen firsthand how integrating AI into my workflow boosts my productivity, whether I am coding or analyzing reports. I've also noticed how other industries use chat-based interfaces to help users find information quickly. It made me ask: why can't we do the same in the energy industry?

The energy sector has been slow to adopt modern software. For example, we are still submitting applications by emailing zipped folders of documents. This process is prone to errors, hard to track, and often hits email file size limits. It quickly became clear to me that this industry needs more software and better automation. Energy demand is growing faster than ever, driven in part by the rapid expansion of data centres. We cannot afford to ignore the role of software and automation any longer.

This realization encouraged me to learn web development and AI engineering. Over the past two years, in my spare time alongside my grid engineering work, I’ve been studying these topics. My background in programming helped me understand the tools and frameworks. LLMs were a huge help, making it easier to learn new concepts and find reliable information. They act as an accelerator, making it possible for anyone to pick up new skills reasonably quickly. Once I gained enough confidence, I started building Grid SensAI as a personal project.

What is Grid SensAI?

Grid SensAI started with a simple concept: enabling users to ask questions about the grid data in plain English and receive useful answers.

Rather than manually filtering spreadsheets or writing complex queries, a user can simply input a request such as:

"Show me all solar projects in Yorkshire above 10 MW that are connected"

The system handles the rest. It translates the question into structured filters, executes the query, and presents the results in a table, map, or chart, depending on what best suits the analysis.

A key design decision was ensuring the LLM does not have direct access to the database, as this would pose security and reliability risks. Instead, the application uses a custom filter schema designed to cover a wide range of potential queries. The LLM populates this schema, and the backend processes it. The guiding principle is to use LLMs only where it genuinely adds value which is interpreting natural language and mapping it to intent. For the rest, the app relies on straightforward, deterministic code for query execution and data handling, so the system stays predictable, testable, and auditable. I will cover the specific technical details of the schema and the broader system design, as well as the pros and cons of this approach, in much greater depth in upcoming blog posts

When a query is run, the results become an "artifact" - a saved view tied to the specific conversation. The data can be explored as a table, visualized on a map, or analyzed through charts showing technology breakdowns and capacity distributions.

The Core Tech Stack

I want to briefly mention the tools being used, though I will save the deep dives for future posts:

  • TanStack Start: A React-based full-stack framework handling routing and server-side rendering (and generation). It is still relatively new, but the developer experience is solid.

  • Convex: This serves as the backend and database. Convex handles real-time sync, serverless functions, and data storage. The agent framework from Convex made building the conversational layer much easier.

  • Shadcn - Tailwind: Honestly, there's not much to say about this choice. It has quickly become the industry standard.

  • Clerk: Used for authentication so each user has their own account and data space. It is simple and reliable.

  • OpenRouter: We route our LLM calls through OpenRouter. Currently, we are using Kimi K2, an open source LLM hosted on Google Cloud. This gives the application the flexibility to switch models without changing code.

  • Mapbox: Used for map visualizations. It is fast, customizable, and handles thousands of points without breaking a sweat.

The Vision

Right now, Grid SensAI focuses on NPG's Embedded Capacity Register. However, this is just the start.

The goal is to bring together grid data from across the UK, covering different DNO regions and datasets, into one place. Once a unified view is established, several possibilities open up:

  • Cross-regional analysis

  • Network-wide capacity planning and analysis

  • Potentially supporting elements of power system studies by working with CIM models

The architecture might evolve as more data sources are added and user needs become clearer. I am building this iteratively, starting with a focused use case and expanding from there.

What's Next?

In upcoming posts, I will cover:

  • Why I chose TanStack Start over Next.js

  • How Convex's agent framework powers the conversational AI

  • The design decisions behind the filter schema

  • And more…

If you are working with UK grid data and want a better way to explore it, give Grid SensAI a try. I’d love to hear any feedback or feature requests as I continue developing this in my own time. I’m also open to pairing with others interested in the intersection of power systems and modern web/AI tooling.