Natural-Language Analytics Platform
- Databricks
- Genie
- SQL
- Python
- →Phased rollout toward roughly 1,400 staff; around 200 trained so far
- →Close to 10 production tables built as the analytical layer
- →Genie instruction layer constrained to answer only from verified data
Context
The institution is rolling out a self-serve analytics platform so academic and corporate staff can ask questions of institutional data in plain English instead of writing SQL. It is built on Databricks Genie Spaces, aimed at an institution-wide rollout of roughly 1,400 staff by the end of 2026, and currently live in a phased pilot: around 200 people trained and onboarded with access to curated datasets, while data engineering continues piping the remaining source systems in ahead of the wider rollout.
My part: the analytical layer
My work sits between the raw imported data and the people asking questions. I built close to ten production tables from the raw feeds the data engineer supplies, shaped specifically to support live demonstrations and everyday queries on the platform. This is the transformation stage of the ETL: taking ingested data and reshaping it into a structure Genie can query reliably and return accurate answers from.
Grounding the model
A natural-language layer is only as trustworthy as its guardrails. I configured Genie's instruction layer with enough institutional context, school structures, column definitions, local terminology, for it to interpret staff questions correctly, and constrained it to answer strictly from verified Databricks data so it does not fill gaps with plausible-sounding fabrication.
Adoption is the real challenge
The platform has shipped to real users and departments are asking for more of their data to be loaded so they can build their own use cases. A formal impact measure is not yet in place given how early the rollout is. The hard part has been adoption: some staff are proactively engaged, others were assigned the tool with less motivation, and most have little coding background, which shapes how both the platform and the training have to be designed to stay accessible rather than technical. I run the hands-on training sessions for users who want to go further.