Case study 01 — Data
BuildingMinds — Download, create, ask: a ladder to your data
BuildingMinds builds a "digital brain" for real estate portfolios. The platform collects data from buildings — energy meters, invoices, certificates — and turns it into one clean data layer that powers ESG reporting and CO₂ reduction planning.
The typical user is an ESG Manager or Asset Manager at a fund managing hundreds of buildings.
Role
Product research, UX & UI
Timeline
2025 — 2026
Scope
Dataset Downloader, Dataset Creator & AI Chat — web platform
The problem
One big table behind many graphics. Each dashboard page was powered by a single large data table. Fine for a broad overview — but any detail question meant digging into and reshaping that table by hand.
Every dataset was a service request. Custom datasets went through the Professional Services team. One request, effort three times over — customer, Professional Services, and the teams behind them.
Platform filters answered some ad-hoc questions, but nowhere near as flexibly as users wanted. Creating a personalized mix of their own data simply wasn't possible.
The vision
Instead of designing one big feature, we planned the work in versions that build on each other — each version gives the user real functionality early, while giving the developers a realistic time frame to achieve the work:
V1 — TAKE - Dataset Downloader - Download the most requested datasets
V2 — MAKE - Dataset Creator - Build your own dataset from all available KPIs
V3 — ASK - Proxima AI Chat - Ask questions about datasets, or create them by prompting
This staging was also a de-risking strategy: the Downloader validated demand for self-service data access, the Creator validated the KPI model, and both created the structured foundation the AI would later need to give trustworthy answers.
The approach
V1 — Dataset Downloader
For V1, we turned the three most commonly requested datasets into self-service downloads — always updating according to the user's portfolio.
Design challenge
Downloading a file is trivial — the challenge was everything around it. The flow had to feel intuitive and flawless, built entirely with existing components of our design system, while still giving each dataset a clear description so users get context to their data: what each column means, how it was calculated, and how up to date it is.
Key decisions
A flow reduced to the essentials. The overview lists all available datasets with a short description each, and users can multi-select the ones they need and download them in one go — no detours, no extra steps.
Context on demand. For users who want to understand the data before working with it, each dataset has a detail page with a full explanation of the dataset and a glossary describing every single column.
V2 — Dataset Creator
From "take what we give you" to "build exactly what you need."
As the second iteration, the Dataset Creator lets users choose from all available KPIs to create and download their own dataset.
Design challenge
The Creator exposes the platform's full KPI catalogue. The danger was obvious: a wall of 100+ KPIs, each with technical configuration options (accuracy types, normalization, emission scenarios…), could easily become an expert-only tool. My goal was to make composing a dataset feel like a guided flow, not like writing a database query.
Key decisions
Progressive disclosure in three layers. I structured the complexity so users only meet it when they need it:
KPI source first — users pick a data source and see a searchable, checkbox-based KPI list.
Configuration after selection — every KPI is editable. After selecting, users can fine-tune each KPI through the "Configure Dataset Fields" dialog (e.g. energy consumption by accuracy type). Each KPI card shows a quiet hint like "7 possible configurations", so the depth is discoverable without being in the way.
Filters at creation time — filters (time period, data normalization, plausibility, emission reporting type, CRREM scenario, retrofit scenario) appear only after clicking Create, and only the filters relevant to the selected KPIs are shown.
V3 — Proxima AI Chat
The dataset creator becomes a conversation.
Design challenge
With the structured Dataset creator in place, the natural next step: let users simply ask. In the Proxima AI Chat, users can ask questions about the existing datasets — and, through prompting, let Proxima create a new dataset for them. "Show me the buildings with the biggest consumption increase in 2023" should not require picking KPIs and filters manually.
Research & exploration
I analyzed how leading AI products solve trust and data interaction — in particular Perplexity (source citations, transparent steps, follow-up suggestions) and Microsoft's Data Formulator (natural-language chart creation combined with a visible, editable field/chart configuration).
The key pattern I took from both: AI output must stay inspectable. Perplexity shows sources; Data Formulator shows the concrete field mapping behind every AI-generated chart, with the warning "AI generated results can be inaccurate, inspect it!". For ESG data that goes into regulatory reports, this transparency is not optional.
Concept
I explored embedding a Data-Formulator-style workspace directly into the platform: the user asks a question in natural language, the AI builds the dataset/chart, and the underlying data table plus the field configuration stay fully visible and editable. Suggested prompts per dataset ("Plot monthly consumption to identify seasonal trends") lower the entry barrier for non-analysts.
This connected to the platform's broader AI direction — the "Ask me anything" assistant entry point in Proxima:
Why the versions mattered here: the AI chat is only trustworthy because V1 and V2 exist. Every AI answer can resolve to a concrete, inspectable dataset — the same datasets users already know from the Downloader and Creator. The AI doesn't replace the manual tools; it's a faster door into them.