Agentic AI Demos: Snowflake Cortex Code and Precisely's MCP Server

Watch the full session · 36 min

At Data Science Connect's Q3 Agentic AI Showcase, Snowflake's Tyler Richards built a dashboard and Slack alerts from written specs using Cortex Code, and Precisely's Carmen Adame had an AI assistant assemble a property risk profile through an MCP server.

Two vendors used live, unscripted demos to show what agentic AI looks like in day-to-day data work at Data Science Connect's Q3 Agentic AI Showcase, a virtual event sponsored by Snowflake and Precisely. Tyler Richards, Staff Data Scientist at Snowflake, showed how he builds data products from written specs with an AI coding agent. Carmen Adame, Senior Manager, Product Marketing at Precisely, showed an AI assistant pulling location and risk data through an MCP server.

Each presenter had about 10 minutes, and both worked through live-demo hiccups on screen. Here's what each demonstrated.

Snowflake: What is spec-driven development with AI agents?

Richards works on Snowflake's product data science team. His typical job starts with raw product logs, adds materialized tables to speed up analysis, and ends with a visualization layer. He used to code all of that himself, then started handing it to AI agents.

The problem, he said, was that everything he told the agent lived inside a prompt. Later he couldn't remember why he had built something a certain way. His fix is what he calls spec-driven development. "I write a paragraph or I have AI write a paragraph about what I want to do and why I want to do it," he said. The prompt then simply says to build it.

Because the spec persists, it can be reviewed later or turned into a reusable skill, Richards said. He added that he does far more spec-driven work now than he did six months or a year ago.

How did Snowflake's Cortex Code demo work?

To avoid showing internal data, Richards built a synthetic point-of-sale dataset for an ice cream shop near his home in Seattle, with a scheduled task adding new transactions on a regular cadence. He worked in Cortex Code Desktop, which he described as an IDE built around Cortex Code and Snowflake data.

From a short spec listing the goal, columns, descriptions and a sample row, the agent built a dynamic table that refreshes every 10 minutes. A second spec described a Streamlit dashboard's data sources, layout and sections. Richards said a spec like this is easier for a co-worker to review in a pull request than the generated SQL or Python.

The Streamlit build failed several times on screen. Richards called it "the curse of doing these things live," but the agent kept fixing errors, including mistakes he said were in his own spec, until the app ran with top flavors, an employee leaderboard and sales by hour.

He finished with Cortex Automations, a scheduled product he said can watch the metrics named in a spec, explain why one changed and send a summary to Slack. He showed a daily digest it had sent him with revenue, top employee, peak hour and top flavors.

Asked where customers get the most value, Richards pointed to building data products such as Streamlit apps, agents, dynamic tables and dbt models. The agent "just knows a lot more about the Snowflake ecosystem than other agents," he said, so he didn't have to explain tools like Snowpark Python.

Precisely: How does location intelligence support agentic AI?

Precisely focuses on data integrity, which Adame described as data that is ready for AI and agentic AI. Her demo covered location intelligence and data enrichment, which Precisely makes available through APIs and MCP servers.

Companies already hold a great deal of location data about customers, providers and assets, she said. Enriching it with more context "truly makes the data better."

What did Precisely's MCP demo show?

Adame connected Claude to a Precisely MCP server and asked, in one sentence, for a risk profile of 401 Biscayne Boulevard in Miami and whether it was a good underwriting risk. The assistant wasn't told which data sources to use. It decided on its own to geocode the address, then called tools for flood, coastal and wildfire risk.

The result described the building's mixed commercial use and square footage, placed it in a 100-year flood plain, and rated wildfire risk as minimal. Adame then asked it to profile a well-known corporate campus in California and say which property it would flag first and why.

Asked what MCP access enables that another LLM connection couldn't, Adame said nothing: "This is a demo of the data." Precisely's value is the data itself, she said, delivered however customers want to consume it.

She also showed a dashboard built for a fictional telecom company planning expansion in Texas. It combined Precisely location data with building permit data from Shovels, a partner in Precisely's DataLink program, to map permit activity and carriers around Dallas.

Adame said Precisely has more than 12,000 customers worldwide. As a product marketer rather than a data engineer, she said, "utilizing our LLM tools plus trusted data allows me to build and feel confident about the things that I build."

Catch the next Agentic AI Showcase

The Agentic AI Showcases are one of four series in Data Science Connect's 2027 lineup of 24 free live webinars, alongside the AI × Enterprise Panel Series, AI Demo Days and the Plug & Play Series. See the webinar schedule for upcoming live demos.

Watch: Agentic AI in Action: Live Demos from Snowflake & Precisely | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights