Ten vendors had 10 minutes each at Data Science Connect's GenAI Demo Day Q4. Demos covered agents that answer business questions, semantic layers that keep answers accurate, risk-based AI governance, GenAI for predictive maintenance and observability for agents in production.
Key takeaways
- Agents now explain the why. Snowflake showed an agent combining shipment, weather and contract data to find the cause of delays.
- Semantic layers keep answers accurate. PromptQL and AtScale both used business definitions to correct or translate AI queries.
- Transparency builds trust. Dataiku and PromptQL exposed the SQL, sources and assumptions behind each answer.
- Governance can be risk-based. OneTrust showed AI projects scored against the NIST AI RMF, with sensitive data masked automatically.
- Agents need their own observability. Monte Carlo monitors agent inputs, token use and output quality to catch degradation.
Ten vendors had about 10 minutes each to show working generative AI at Data Science Connect's GenAI Demo Day Q4, a virtual event with headlining sponsors Snowflake and PromptQL. Host Brian Mink, co-founder and president of Data Science Connect, called it the company's most registered and best attended webinar to date.
The demos clustered around a few themes: agents that answer business questions in natural language, semantic layers that keep those answers accurate, and governance and observability for agents in production. Here's what nine of the presenters demonstrated.
Snowflake: Can an AI agent explain why a metric changed?
Haley Massa, Senior AI/ML Architect, Applied Field Engineering at Snowflake, demonstrated Snowflake Intelligence, which she described as an enterprise intelligence agent. Playing a site lead at a fictional manufacturer, she asked for critical inbound shipments. The agent applied the business's definition of "critical" from a semantic layer, and a green shield marked answers the team had certified in advance.
She then asked what was driving delayed shipments. The agent pulled in weather and contract data alongside shipment tables to pinpoint a likely cause. Massa said agents inherit existing Snowflake governance, such as row-based access policies and column masking, so teams don't have to copy controls over.
PromptQL: How can AI admit what it doesn't know?
Rob Dominguez, Engineering Manager at PromptQL, said AI tools are "incredibly confident" even when wrong, a bigger problem with enterprise jargon. PromptQL's approach brings subject matter experts into the conversation when the agent is unsure, then saves their input to the semantic layer. "We see a problem once, we fix it once," he said.
In his demo, a gross margin answer came with a confidence analysis listing its assumptions. Two were supported by the company's wiki and five were not, including the margin formula. Dominguez edited the wiki entry, and the next answer used the corrected formula.
Google Cloud: How do you build an agent with ADK?
Kanchana Patala, Technical Solutions Manager at Google, leads its agents portfolio. She described an agent as a model (the brain), tools, orchestration (the nervous system) and a runtime. She introduced Google's open-source Agent Development Kit, which she said includes a web interface for local testing and one-command deployment to Cloud Run, Kubernetes or Agent Engine on Vertex AI.
She showed open-source samples in Agent Garden, including a RAG agent, and a deployed agent's sessions, memory, evaluations and event traces for debugging.
AtScale: Why pair a semantic layer with an LLM?
"A semantic layer coupled with Gen AI is magic," said Dave Mariani, CTO and founder of AtScale. He described the Model Context Protocol as "like a JDBC for LLMs," letting Claude read AtScale's semantic models and send simple queries that the semantic engine turns into physical database queries.
Asked to "tell me something I don't know about sales," Claude ran and analyzed its own series of queries, then built charts. Mariani showed that a simple inbound query became a Databricks query with the right joins, using an aggregate table to return results in under 55 milliseconds. Because the semantic engine does the translation, he said, results are deterministic.
Dataiku: Can agents take ad hoc questions off analysts?
Ishani Rudra, a sales engineer at Dataiku, used a fictional denim brand whose marketing analyst was buried in ad hoc campaign questions. In Dataiku's Agent Hub, a no-code campaign agent answered which campaign had the highest ROI, looked up definitions in a company glossary and built charts.
Rudra said many LLM tools are opaque: "These are blackbox approaches and we're just hoping for the right response." Dataiku logs each step, she said, showing the glossary entry used, the SQL run on Snowflake and the agent's reasoning in a traces explorer, and agents inherit project permissions, versioning and lineage.
Marketing Evolution: What is synthetic consumer-level ad data?
Winston Bradley, COO of Marketing Evolution, said many ad channels don't provide identity-resolved, consumer-level data. Its MEVO platform maps aggregate marketing data to an ontology and uses proprietary algorithms to generate synthetic, timestamped consumer-level exposure data. Customers can use it in their own mix models, attribution models or LLMs, he said. He showed a scenario planner that optimized a $2 million, four-month budget down to the placement level.
OneTrust: How do you triage AI projects by risk?
Victor Fronza, Senior Solutions Engineer at OneTrust, cited a OneTrust survey of more than 1,250 IT and governance decision makers that found teams spending 37% more time managing AI risk than the year before. He said most organizations are in the first two of four AI governance maturity stages.
His demo assessed a facial recognition project for office entry against the NIST AI Risk Management Framework. It was flagged very high risk, OneTrust detected exposed Social Security numbers and triggered a masking policy in Snowflake, and the system generated a model card as an auditable record. Low-risk projects can be approved automatically, he said, while medium- and high-risk ones go to experts.
SAS: Can GenAI turn an anomaly alert into a repair plan?
Daniel Dasari, a senior manager in SAS's IoT division, said predictive maintenance often flags a problem without its root cause. SAS combines its machine learning anomaly detection on streaming sensor data with its retrieval agent manager, a no-code RAG tool built on manuals, maintenance logs and user forums. The result, shown on a dashboard for an industrial fan, was an alert with a probable cause, recommended actions and source references. He said the tool can run fully on premises.
Monte Carlo: How do you monitor an agent in production?
"Everyone knows bad data is going to lead to bad AI outputs," said Ethan Post, Field CTO at Monte Carlo. He said Monte Carlo runs three agents in production across more than 30,000 users. Its troubleshooting agent traced a real freshness alert to a Fivetran replication delay, which turned out to be linked to the previous day's Cloudflare outage.
Post said agent observability covers inputs, operational metrics such as runtime and token use, and outputs. Monte Carlo's instrumentation library collects traces and spans into an inference table, where LLM-as-a-judge evaluations score outputs. In the demo, a drop in an invoice collection agent's completion score revealed overly short emails.
See the next AI Demo Day
AI Demo Days are one of four series in Data Science Connect's 2027 lineup of 24 free live webinars, alongside the AI × Enterprise Panel Series, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming demo sessions.
Frequently asked questions
What was shown at GenAI Demo Day Q4?
Ten vendors, including Snowflake, PromptQL, Google, AtScale, Dataiku, Marketing Evolution, OneTrust, SAS and Monte Carlo, gave 10-minute demos of generative AI and agent tools for the enterprise.
How does a semantic layer help LLMs answer data questions?
AtScale's Dave Mariani showed an LLM sending simple queries through MCP to a semantic layer, which translated them into physical database queries with the right joins so results are deterministic.
What is agent observability?
As Monte Carlo's Ethan Post described it, monitoring an AI agent's data inputs, operational metrics such as runtime and token use, and output quality, using traces, spans and LLM-as-a-judge evaluations.
How can GenAI help predictive maintenance?
SAS combined machine learning anomaly detection with a RAG tool built on equipment manuals and maintenance logs, so alerts come with a probable root cause and recommended actions.
Watch: GenAI Demo Day Q4 on YouTube
Event: Webinars · More coverage: Webinars insights
