Four vendors had about 10 minutes each to demo working products at Data Science Connect's virtual GenAI Demo Day. Rafay showed self-service GPU access, AnswerRocket a conversational analytics assistant, Crowdbotics AI-generated product requirements and code, and DataRobot prebuilt AI apps with real-time guardrails.
Key takeaways
- Shared GPUs, self-service access. Rafay showed developers launching inference endpoints and fine-tuning jobs from a portal without requesting GPUs directly.
- Breaking the analyst bottleneck. AnswerRocket's Max answered vague business questions with trend and market share analyses.
- AI for the software life cycle. Crowdbotics generated product requirements from an idea or an existing repository, then starter code.
- Prebuilt apps, governed. DataRobot demonstrated forecasting and AI analyst apps with guardrails that blocked a jailbreak attempt in real time.
- Agents choose their tools. DataRobot's analyst agent ran SQL, then decided to use Python to build a chart.
Four vendors had about 10 minutes each to show working generative AI products at Data Science Connect's GenAI Demo Day, a virtual event with partners Rafay, AnswerRocket, Crowdbotics and DataRobot. Brian Mink, co-founder and president of Data Science Connect, hosted and said the event drew hundreds of attendees from around the world.
The demos covered the generative AI stack from the bottom up: GPU infrastructure for developers, natural-language analytics for business users, AI-assisted software planning and prebuilt AI applications with guardrails. Here's what each company showed.
Rafay: How can IT give developers self-service access to GPUs?
Mohan Atreya, Chief Product Officer at Rafay, described two frustrated groups. Developers and data scientists lack self-service access to compute and data, and environments that work for one person can't be reproduced by a colleague. IT teams spend millions of dollars on GPUs that sit underused. When IT can't keep up, developers "go rogue," he said, swiping a credit card for cloud GPUs and creating shadow IT and duplicate environments.
Atreya's demo started from an IT administrator's view of a shared cluster with 12 GPUs. A developer then logged into a self-service portal, chose from a list of company-approved models and launched an inference endpoint without requesting GPUs directly. He ran a simple Python script from his laptop against a model running in a data center in Phoenix, with no VPN.
Fine-tuning followed the same pattern. The developer picks a pre-approved data set location, a base model and a place to store the result, then serves the fine-tuned model from a new endpoint. Atreya said customers have teams going live with fine-tuning and inference in hours.
AnswerRocket: Can business users get analysis without waiting on analysts?
Joey Gasparic, an account executive at AnswerRocket, described what he calls "the vicious cycle." A business user sees a problem in a dashboard, asks an analyst, waits hours or days, then has 10 more questions. Many users stop asking and make educated guesses instead, he said.
AnswerRocket's assistant, Max, connects to data where it lives. In the demo, Gasparic used Nielsen pasta sales data in an Azure Synapse data warehouse. A vague question, "How is pasta doing?", produced a trend analysis with generated insights. A follow-up broke the trend out by brand, and asking why one brand was up triggered a market share analysis instead.
He then asked Max to research the top challenges facing the pasta market. That question ran against an unstructured industry article, with a citation. Finally, Max summarized the thread as an email and translated it into Spanish. Gasparic said AnswerRocket also helps companies building their own generative AI analytics for structured data.
Crowdbotics: How can AI speed up the software development life cycle?
Jason, an Enterprise and Strategic Solutions Architect at Crowdbotics, said the company started in app development and used lessons from building 25,000 apps to create a platform that automates the software development life cycle.
His demo began with "idea to spec." A user enters a short prompt plus business context, approved vendors and architecture preferences, and the platform generates a product requirements document organized into categories, features, descriptions and acceptance criteria. Users can ask AI to suggest features, analyze security risks, comment on items and edit them. A "code to spec" option works in reverse, generating requirements from an existing GitHub repository, such as an old COBOL mainframe system being modernized.
From the requirements, the platform analyzed a feature and generated about 20 starter code files. He also showed a Crowdbotics extension for GitHub Copilot in VS Code that brings the requirements document's context into Copilot chat, so engineers can ask about features and generate code.
DataRobot: What do prebuilt AI apps with guardrails look like?
Justin Swansburg, VP of Applied AI and Data Science at DataRobot, skipped slides and went straight to DataRobot's app gallery of customizable templates. He said the hard part of a solution such as a forecasting assistant is stitching together the forecast model, web app, deployment and a language model that explains results.
His demand forecasting app, connected to Snowflake, forecast fresh produce sales with confidence bounds and plain-language explanations of what drove each forecast up or down. A scenario planner showed the expected effect of running display ads for apples.
An AI data analyst app used what Swansburg called an agentic workflow. Multiple language models read the schemas of connected orders and products tables, generated descriptions and added them to the system prompt. Asked to summarize order volume, the agent ran SQL, then chose Python to build a chart. Users can inspect each step.
He closed with governance. DataRobot's guardrails moderate topics, PII disclosure and jailbreak attempts, and a console tracks how people use each app. He showed his own earlier attempt to make a model reveal its system prompt, which was blocked in real time.
Catch the next Data Science Connect demo day
Demo days are part of Data Science Connect's 2027 schedule of 24 free live webinars, alongside the AI × Enterprise Panel Series, Agentic AI Showcases and Plug & Play Series. See the full lineup on the webinars page. To see enterprise AI leaders in person, COLLIDE 2027 takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center in Atlanta.
Frequently asked questions
What is self-service GPU infrastructure?
In Rafay's demo, IT pools GPUs in a shared cluster and developers request inference endpoints, fine-tuning jobs or workbenches through a portal, receiving a slice of the shared resources.
How can generative AI help business users analyze data?
AnswerRocket's Max lets users ask questions in plain language and runs trend, market share and research analyses against structured and unstructured data.
Can AI write product requirements documents?
Crowdbotics showed its platform generating a requirements document with features and acceptance criteria from a short prompt or an existing GitHub repository.
What are AI guardrails?
In DataRobot's demo, safety checks that keep models on topic, prevent PII disclosure and block jailbreak attempts, with usage monitored in a console.
Watch: GenAI Demo Day on YouTube
Event: Webinars · More coverage: Webinars insights
