At Data Science Connect's GenAI Demo Day Q2, seven vendors had about 10 minutes each to show working products. Demos ranged from Snowflake Cortex Agents and Gemini 2.5 on Vertex AI to self-service AI analytics, contextual location data, data pipelines and agents that write and debug code.
Key takeaways
- Agents need both data types. Snowflake's Cortex Agents route questions to SQL generation for structured data and search for unstructured documents.
- Context windows keep growing. Skinner said Gemini 2.5 Pro accepts 2 million tokens and lets developers set a thinking budget.
- Context is king. Precisely's Myung said linking data sets, such as by location, makes them far more useful to AI models.
- Trust comes from a semantic layer. Janssen said business users trust AI answers they can read as measures and dimensions, not SQL.
- Agents are entering data science. Dataiku and Zerve showed prompt comparison, automated alerts and an agent that fixed its own coding error.
Seven vendors showed how generative AI is being applied to data access, data pipelines and analytics at Data Science Connect's GenAI Demo Day Q2, a virtual event headlined by Snowflake and made possible by Google Cloud, Dataiku, Precisely, Zenlytic, Zerve and Fivetran. Each presenter had about 10 minutes for a live demo. Brian Mink, co-founder and president of Data Science Connect, hosted.
The presenters were James Cha-Earley, Senior Developer Advocate at Snowflake; Holt Skinner, Developer Advocate, Cloud AI, at Google; Nohyun Myung, Senior Director, Product Management, at Precisely; Ryan Janssen, CEO and Co-Founder of Zenlytic; Brad Moore, Senior Sales Engineer, Enterprise, at Fivetran; Valentina Peña-Vivas, Senior Data Scientist at Dataiku; and Greg Michaelson, Co-Founder and Chief Product Officer of Zerve. Here's what each demonstrated.
Snowflake: How do Cortex Agents combine structured and unstructured data?
Cha-Earley introduced Snowflake Cortex Agents, then in public preview. The agents use two tools: Cortex Analyst, which generates SQL against structured data using semantic models, and Cortex Search, which retrieves information from unstructured documents. The agent plans, splits a request into smaller tasks, routes them to the right tool and reflects on interim results, he said. Agents are available through REST APIs.
In the demo, a fictional insurance company asked how many claims were open and how many exceeded $10,000. Cortex Analyst returned the SQL and results, and a new data-to-charts feature produced a visualization. A question about rental car coverage returned an answer with citations to the source documents. Semantic models let teams add synonyms, so the system knows that "some people may say claims, some other folks may call them like invoices," Cha-Earley said.
Google Cloud: What can developers build with Gemini 2.5 on Vertex AI?
Skinner walked through Gemini 2.0 and 2.5 on Vertex AI, which he said is the option for enterprise-grade security and compliance, while the Gemini Developer API suits experimental and lighter production use. He highlighted the large context window: 2 million tokens for Gemini 2.5 Pro and 1 million for other versions.
His notebook demo, from the Vertex AI cookbook, covered native reasoning with an adjustable thinking budget, structured output for loading results into databases, function calling, code execution and grounding with Google Search, which returns citations instead of making up an answer. On the industry buzzword, he joked: "I really hate the word agentic."
Precisely: Why does AI need contextual data?
Myung said hallucinations and bias often trace back to a lack of data accuracy, quality and precision. He recalled that at a previous job, 70% to 80% of his data science team's time went to data engineering. "Context is king in data," he said, meaning data sets must link to one another to be useful.
He demonstrated the Precisely Data Graph API, a GraphQL-based interface to hundreds of data products that share location attributes. Querying a Denver address returned the building, its parcel and the businesses inside it. Drawing a boundary around a Las Vegas neighborhood returned counts of buildings, parcels and businesses, the kind of context someone would need to evaluate a site for a new coffee shop.
Zenlytic: How do self-service AI analytics build trust?
Janssen ran an unscripted demo of Zoe, Zenlytic's AI data analyst, connected to Snowflake and using Gemini 2.5 Pro. Asked by voice for year-to-date revenue by product line, Zoe answered in business measures and dimensions instead of SQL. "The vast majority of people who can benefit from data don't speak SQL," Janssen said. Without a semantic trust layer, he said, users simply email the data team to ask whether a number is right.
He then showed Zoe handling an ambiguous request from a new marketing manager who wanted a summary of paid channel efficiency, and running a clustering analysis in Python to identify five customer archetypes.
Fivetran: How do teams get data ready for AI?
Moore said many customers are stuck experimenting with AI tools and recommended starting with the business outcome instead. Fivetran moves data from more than 700 sources and keeps it fresh with incremental updates, he said, and offers a managed data lake service on Amazon S3, OneLake and Google Cloud. He said Fivetran had just acquired Census, which adds the ability to write data back to applications such as Salesforce.
His demo set up a SQL Server destination and a Salesforce connector, picked a table, ran the initial sync and showed the 16 sandbox records arriving, with schema changes handled automatically.
Dataiku and Zerve: How do AI platforms speed up data science work?
Peña-Vivas played a data scientist at an ice cream parlor tracking a trending flavor on social media, with every step built point-and-click. She used Dataiku's LLM recipes for entity extraction, Prompt Studios to compare models on accuracy and estimated cost, an automated scenario that sends Slack alerts when trends spike, and Data Stories, which builds slide decks with live data and creates charts from plain-English requests.
Michaelson, whose first attempt was cut short by screen-sharing problems, returned at the end to show Zerve's agent. Given a CSV file, the agent proposed a plan, wrote and ran code blocks, hit an error, fixed it on its own and trained a probability-of-default model. Each block runs in its own kernel, he said, so several people can edit and run code at the same time.
Upcoming AI Demo Days and webinars
AI Demo Days are one of four series on Data Science Connect's 2027 calendar of 24 free live webinars, alongside the AI × Enterprise Panel Series, Agentic AI Showcases and the Plug & Play Series. Each Demo Day gives vendors short slots to show working products. See the webinar schedule.
Frequently asked questions
What are Snowflake Cortex Agents?
According to Cha-Earley, they are agents that use Cortex Analyst for SQL on structured data and Cortex Search for unstructured documents, available through REST APIs.
What is the context window of Gemini 2.5 Pro?
Skinner said Gemini 2.5 Pro can take 2 million tokens of input, while other Gemini versions take 1 million.
Why do self-service AI analytics tools use a semantic layer?
Janssen said most business users don't read SQL, so presenting results as defined measures and dimensions lets them understand and trust the answer.
How do you choose an LLM for a data pipeline task?
Peña-Vivas showed Dataiku's Prompt Studios, which compare models side by side on test cases for accuracy and estimated cost.
Watch: GenAI Demo Day Q2 on YouTube
Event: Webinars · More coverage: Webinars insights
