Back to Tech BlogTechnology
June 19, 2025
Neptune Days at TransitionZero: two days of innovation and prototyping
Summary
Neptune Days are TransitionZero’s take on hackathons — two days set aside for team members to explore creative, cross-functional ideas that support our mission, outside of the core roadmap.
This round saw a strong focus on applying cutting-edge AI techniques to real-world energy and climate challenges. — from AI-powered chat interfaces for energy modelling and natural language scenario creation, to machine learning for detecting cement kilns via satellite imagery. Teams also prototyped tools for visualising dispatch results, centralised data pipelines for Mixpanel and HubSpot, a helpful internal go/links system, a data chatbot for querying the warehouse, and even an early benchmark to test energy-modelling AIs — all pushing forward innovation at TransitionZero.
The results show that with focused time and collaboration, prototypes can be brought to life remarkably quickly, and may even go on to be included in our product roadmap.
What are Neptune Days?
Every so often at TransitionZero, we press pause on our regular work to make space for something a little different: Neptune Days. These two-day innovation sprints are our take on hackathons: two days set aside for experimentation, collaboration, and rapid prototyping outside the constraints of our core roadmap. Named after the planet, which in astrology represents creativity and inspiration, Neptune Days are designed to let our curiosity and imagination take the lead, giving the team space to test new ideas and technologies across our internal tools and existing products.
This Neptune Days event saw a strong focus on applying cutting-edge AI techniques to real-world energy and climate challenges. From natural language scenario-building and AI-powered data assistants, to satellite image analysis and automated modelling insights, our team explored how advanced models like Gemini and GPT-4o can enhance our products, improve accessibility, and accelerate impact. These prototypes not only showcase what’s possible in 48 hours — they also signal where TransitionZero is headed: toward smarter, faster, more integrated tools to support the global energy transition.
The projects
In June 2025, TransitionZero team members narrowed down over a dozen ideas to a handful of workable projects, which were then rapidly prototyped and presented to the rest of the company.
Making energy modelling more conversational with AI
Lowering the barriers to energy systems modelling is key to speeding up the transition to sustainable energy. This project explored how large language models (LLMs) can make this easier by letting analysts interact with Scenario Builder using natural language.
Using Vercel’s AI SDK and OpenAI’s GPT-4o, combined with deep reasoning from Google Gemini 2.5 Pro, the team integrated Scenario Builder’s API endpoints as tools the LLM can call. This means analysts can create scenarios, trigger runs, and explore detailed insights just by chatting with the system.
Next steps include refining how the model interprets inputs to guide feasibility, helping users configure scenarios aligned with their research questions, and strengthening security and performance. Deeper AI integration is on the horizon — making energy modelling more accessible and interactive than ever.
Estimating real-time solar generation with TZ-SAM
This project idea builds directly upon an existing dataset: TransitionZero’s Solar Asset Mapper (TZ-SAM). What if we could use weather forecast data to predict solar generation from TZ-SAM assets?
The team approached this question by focussing on five test-case solar plants. They pulled weather forecast data from open-meteo.com, which they fed into a python package (atlite library) to predict solar generation. They benchmarked this result against historical generation data to get a sense of their model accuracy. A second component of the project was to create a slick dashboard to display the results.
After two days they successfully built a streamlit app displaying quite accurate and insightful results for the five test-case solar plants, but the team concluded a lot of work would be required to match the full TZ-SAM dataset with historical data and ‘messy’ weather data.
Rapidly prototyping dispatch results with AI
Adding economic dispatch modelling to Scenario Builder is a key priority for the team at TransitionZero, and to do that we need a way to show dispatch results clearly and flexibly. One project tackled this by building a high-fidelity prototype for displaying hourly generation outputs, which can be easily aggregated at different time resolutions.
This project was developed by our Scenario Builder Product Manager, who leveraged AI tools to create a high-fidelity prototype. In their own words, their limited understanding of the existing codebase means that this prototype probably won’t be directly reused in the final build. But the fact that a non-coding team member was able to deliver a usable prototype in just one day was a big success.
Next steps are to collaborate with the analysis team to refine the prototype, then build the real version into Scenario Builder. It’s a promising start for a critical feature, proving that innovation can come from all corners of TZ.
An AI benchmark for energy modelling
What if we had a standard test to measure how well AI understands — and can reason about — energy systems? Inspired by Humanity’s Last Exam, a rigorous benchmark designed to challenge large language models across academic disciplines, this project set out to create a similar benchmark for energy modelling.
The team began developing a set of high-quality, nontrivial, and verifiable questions spanning areas like energy systems knowledge and model debugging in code. These aren’t the kind of prompts you can solve with a quick web search — they’re designed to probe real reasoning and domain understanding.
The different AI models had mixed results when tested against the 20 questions drafted so far. However, more questions need to be developed to robustly compare model performance. Next steps include expanding the question set, testing it on a variety of external models, and benchmarking internal GenAI prototypes.
Smarter shortcuts with go/links
This next project looked at making it easier to navigate the maze of links we all use at TransitionZero. Right now, finding the right document, dashboard, or tool often means digging through bookmarks or clicking around in shared spaces. Enter go/links — short, memorable keywords that redirect to long URLs when typed directly into a browser (e.g. go/roadmap).
The team explored setting up go/links using a Chrome extension that lets users create custom shortcuts tied to their TZ accounts. Multiple URLs can even be grouped under a single go/link. The result? A handful of working examples already in use — with potential to expand across the organisation.
The next step is deciding whether to adopt it more widely, as its real value depends on uptake — go/links work best if everyone’s on board.
Meet Beep Bop — your (mostly) helpful Data Chatbot
Getting answers from our data warehouse or platform database usually requires our team members tunnelling in via SSH, knowing the schema by heart, and writing the perfect SQL query. Or asking their colleagues in Slack and waiting for help. This project set out to change that, by building Beep Bop, a chatbot that (in theory) lets you ask data questions in plain English and get a query-ready response.
Built with Streamlit and deployed on GCP, Beep Bop uses Google’s Gemini 2.0 Flash model via Vertex AI. It automatically pulls schema information from both BigQuery and Postgres to help answer questions — though results can vary. Sometimes, it impresses. Other times… not so much.
Still, it's a promising start toward making data more accessible across TZ. Next steps include improving prompt handling, enriching table and column descriptions, and eventually enabling more complex conversations.
Centralising user data for better insights
As our tools — especially Scenario Builder — continue to roll out externally, understanding how they’re being used becomes increasingly important. This was tackled head-on in a project that explored how to bring together user analytics, customer relationship data, and energy data into one place: our data warehouse (DWH). The goal? Enable richer cross-data analysis and better data visualisation so we can better understand and support our users.
The team managed to quickly set up data pipelines to store user analytics data, and showcase the power of visualising it. Handling personally identifiable information (PII) was also a key focus, with recommendations to use GCP-native tools for protection. In the end, this project laid a strong foundation for a secure and unified approach to data at TZ.
Spotting cement kilns from space
Understanding cement emissions at a granular level — critical to our work with ClimateTRACE — means knowing where the kilns are. Not just which facility they're in, but where exactly they sit within those plants and how intensively they’re used. This project explored whether we can use machine learning and freely available Sentinel-2 satellite imagery to detect new kilns in existing plant extensions and newly constructed plants.
Using a modified version of our TZ-SAM workflow, the team began training a model to identify kilns across a dataset of ~2,000 cement plants and ~3,000 hand-labelled kilns. Kilns are significantly smaller than solar installations, so the team is pushing the limits of Sentinel-2 resolution. Early model outputs are showing promise.
It's still early days: no full training run has been completed yet, and the resolution barrier may require future access to higher-res imagery (at a cost). But with just a day or so more work, this project demonstrates how TransitionZero can leverage computer vision, machine learning, and AI capabilities for high-value results. A full proof of concept is within reach — opening the door to more accurate estimates of cement-related emissions.
What’s next?
Neptune Days continue to show that there’s no shortage of innovative, impactful ideas at TransitionZero — and that with focused time and collaboration, prototypes can be brought to life remarkably quickly. Several of the ideas shared here will now be reviewed for potential inclusion in our product roadmap. With plenty of inspiration to build on, the team is already looking forward to the next Neptune Day later this year.
Did any of these projects pique your interest? Get in touch!
