Ask, build, compose: What our 5th Genie Hackathon taught us about Databricks Genie

Ask, build, compose: What our 5th Genie Hackathon taught us about Databricks Genie | Databricks Blog Skip to main content

Summary

Our fifth customer hackathon turned into a hands-on tour of everything Databricks Genie can do, and the ten projects make a great curriculum.

The three tracks line up with the three ways to use Genie: talk to your data with Genie Agents, build with Genie Code, and compose Genie into agents, each aimed at a different kind of user.

From a collections team finally talking to its own data, to an agent that writes a supply-chain brief at 11 PM, the builds show why governed, conversational analytics is becoming a foundation rather than a feature.

How Databricks hackathons work We run these hackathons for a simple reason: the fastest way to learn a product is to build something with it. Each one kicks off with an enablement session where teams get to know the product and where it is headed on the roadmap. Then the building starts. Teams work together for about a week and submit a project, and the strongest solutions earn prizes. This was our fifth, and the product in the spotlight was  Databricks Genie . Why Genie makes for a great hackathon Databricks Genie products let people work with data in plain English instead of SQL, giving it a great range for a hackathon focus. Genie is not one feature; it is a family, and it shows up in three distinct ways: Ask it . A  Genie Agent is a chat interface over your governed data. Build with it .  Genie Code is an autonomous AI partner that assists in data, analytics and ML workflows. Compose it . A Genie Agent can be exposed as a tool that agents and apps call on their own.

Each of those serves different teams across an organization, which is exactly why we ran three tracks. We will follow them in that order, from the business user who just wants an answer to the engineer wiring Genie Agents into a full fleet. The constant underneath all three is that Unity Catalog governs who can see what, no matter how the question is asked, and Genie Ontology provides the shared semantic understanding. Track 1: Talk to your data with Genie Agents Who it’s for : business users who want to ask questions of governed, domain-specific agents . A Genie Agent is a domain-specific chat interface that an analyst curates over a slice of governed data, which can then be shared with business users to ask natural language questions of. The curator can point to Unity Catalog tables, add a few example queries, define the business vocabulary with SQL expressions and metric views, and pin trusted assets (governed functions) for the questions that have to be answered exactly. From there, a business user just types a question and gets back a result, a chart, and the query behind it. What does that look like when real teams deploy Genie Agents? OneTrust ran straight into one of the genuine mechanics of Genie Agents. A single agent is designed to focus on up to 30 tables, which keeps answers fast and accurate, but the data a real analyst cares about at OneTrust spans 190 tables and more than 300 views. So they built a supervisory layer that shards the data across several focused Genie Agents, routes each question to the right one, and stitches the answers back into a single conversation. From the user's seat, nothing changes: they still just ask one agent. Behind the scenes, plain-English self-service now reaches across an entire enterprise estate without giving up the governance that keeps it trustworthy. Another team pointed the Genie Agent at roughly 160K loan records and, crucially, taught it the team's language, defining what a "cure" means and what "DNC" stands for so the model maps everyday questions to the right data. Soon, the collections team could ask in plain English and learn things like the fact that most delinquent loans resolve within about 15 days. The best moment was unplanned: once the vocabulary was in place, the agent started suggesting sharp questions the team had not thought to ask. That is what good context turns a chat box into. Track 2: Build with Genie Code Asking is only the start. The next question is who gets to build, and how fast. Who it’s for: analysts and builders . The semi-technical folks who know their data and can write some SQL, but who used to hit a wall the moment a project needed pipelines, functions, or a polished dashboard. Genie Code is the builder of the family. You describe what you want in plain language, and it does the work: writing metric views, Unity Catalog functions, pipelines, and dashboards, all inside Databricks with no separate dev environment to set up. Because it is deeply integrated with Unity Catalog, it understands your real schema and semantics, so it picks the right joins instead of inventing column names. For an analyst, that is leverage. Work that used to mean a ticket to data engineering or a week of hand-written SQL now takes an afternoon, which is exactly what this track was built to show. One team turned Genie on the data team's own house. They used Genie Code to build a governance intelligence platform that flags dormant reports worth retiring, uses lineage and SQL logic to cluster duplicate reports hiding across the org, and scores whether data is actually ready to be used by AI. It is the kind of cross-cutting governance project that usually needs a quarter and a roadmap. Built with Genie Code, it came together during a hackathon. Procore built an entire analytics experience for a vacation-rental platform without leaving Databricks. Avinash, Abdullah, Amy, and Jason used built-in AI functions like ai_extract() to automatically classify and score listings, then shipped a dashboard of KPIs, year-over-year trends, and forecasts, with a Genie Agent alongside it that answers a portfolio manager's "what amenities should I add to improve satisfaction?" in seconds. A polished, multi-part product, built in days rather than weeks. Fanatics Betting and Gaming built a customer-experience tool that hands managers a ranked, ROI-justified action list on request, end-to-end in an afternoon. Then they did something we loved: they used Genie to stress-test their own churn model, found that two history-based features carried almost all of the signal, and concluded honestly that a simpler approach worked just as well. They even packaged the workflow into a reusable analyst skill. When building is this fast, you can afford to challenge your own work, which is how good...