Six vendors had 10 minutes each at Data Science Connect's AI Demo Day Q1. Mistral AI fine-tuned a model, Google Cloud ran agent-driven churn analysis, Aerospike powered real-time nudges, Precisely added location context, Striim masked data in flight and PromptQL flagged its own assumptions.
Key takeaways
- Fine-tuning can be quick. Mistral's Aaron Delp said everything he showed, including fine-tuning runs, took no more than a few hours.
- Agents can run an analysis end to end. Google Cloud's agents built a pipeline, trained a churn model and traced the cause to one ad creative.
- Agents need state. Aerospike stored checkpoints so a multi-step agent can retry from the last step that succeeded.
- Protect data before it lands. Striim's agents found and masked sensitive fields in flight during an Oracle-to-BigQuery replication.
- Uncertainty beats a wrong answer. PromptQL surfaced its assumptions and turned human corrections into semantic layer updates.
Six vendors had about 10 minutes each to show working AI at Data Science Connect's AI Demo Day Q1, a virtual event featuring Mistral AI and Google Cloud as headlining sponsors, with sponsors Precisely, SAP, Aerospike, PromptQL and Striim. The demos ranged from building and fine-tuning agents to AI-assisted churn analysis, real-time purchase nudges, location context for AI tools, in-flight data masking and an AI analyst that admits what it doesn't know.
Brian Mink, co-founder and president of Data Science Connect, hosted. Presenters were Aaron Delp, Head of Technical Marketing at Mistral AI; Arun Dogar, Senior Product Manager at Google Cloud; Shikhar Suman, Senior Solutions Architect at Aerospike; Ruslan Sultanov, Senior Product Marketing Manager at Precisely; Edward Bell, Senior Director of Global Technical Alliances and Solutions at Striim; and Rob Dominguez, Engineering Manager at PromptQL. Here's what each demonstrated.
Mistral AI: How fast can you build and fine-tune an agent?
Delp worked in Mistral AI Studio, which he described as a platform for agents, workflows, fine-tuning and batch processing. In its playground he showed a customer support agent running Magistral, Mistral's reasoning model, instructed to answer only Mistral and API questions.
He also showed batch jobs for transcription and OCR, noting that enterprises may want to process upward of a million documents at once, and a real-time document AI run on a low-resolution PDF with poor handwriting, marked-up text and checkboxes. For fine-tuning, he took Ministral 8B, an 8 billion parameter model he said is built for on-premises and edge use, and trained it to be afraid of ghosts. He said everything he showed took "no more than a few hours of work," including the fine-tuning runs.
Google Cloud: Can agents run a churn analysis end to end?
Dogar played a credit reporting company whose campaign brought in customers who churned faster than expected. A data engineering agent generated a pipeline from a plain-language prompt, joining three tables and adding a churn column. A data science agent then proposed a step-by-step plan to train an XGBoost model and run a SHAP analysis, fixing a warning along the way. The result pointed to one creative, a free dark web scan, that drew one-time users who then left.
In a second scenario, a consumer packaged goods company saw cash flow drop after moving from net 30 to net 60 payment terms. Using conversational analytics on a connected data set, Dogar asked for cash flow by month, then by segment, then a six-month forecast and days outstanding. The problem traced to the hospitality segment, not the payment terms. Work that might take a team "a couple of days," he said, "can be done within a few minutes."
Aerospike: How do AI agents nudge shoppers in real time?
Suman built an e-commerce app in which adding items to a cart triggers an agent. It calls a classic machine learning model to score churn and purchase propensity from features stored in Aerospike, uses a reasoning model to decide whether a nudge is needed, then has a language model write a personalized message. In the demo it sent an in-app message and called a webhook to create a coupon.
Aerospike also served as the agent's checkpoint store through LangGraph, saving state at each step so a failed run can retry from the previous checkpoint. Suman said a single user interaction can fan out to hundreds of database calls, which is why he pointed to Aerospike's hybrid memory architecture, which keeps indexes in RAM and data on SSDs, for consistent low latency at scale.
Precisely: How can AI tools use location context?
Sultanov said that without trusted location context, a model in the best case gives no answer, and in the worst case "it will simply invent an answer." He cited figures that 70% of time goes into cleaning and preparing external data and that only 12% of organizations consider their data AI-ready.
Working in Claude Desktop, he used Precisely's MCP server to complete and geocode an address, then asked for more about the location. The server returned property details and nearby places linked through a Precisely ID. A second demo combined that context with consumer spend data in a site selection report. He said the MCP server is free to use with a Precisely key and secret.
Striim: How do you protect sensitive data in a real-time pipeline?
Bell built an Oracle-to-BigQuery pipeline with drag-and-drop source and target components, based on log-based change data capture. Before replication, Striim's Sherlock agent scanned the source table and flagged sensitive data identifiers. Its Sentinel agent then masked phone numbers and Social Security numbers in memory, before the data landed in BigQuery.
He also showed monitoring and alerting, and cited a joint customer, VMO2, which he said uses Striim and Google for proactive network intelligence, and mentioned Validata, a new product for checking that target data matches the source.
PromptQL: What should an AI analyst do when it isn't sure?
Dominguez opened with a question: "What's worse? An AI tool that says I don't know or one that gives you the wrong answer?" Asked for gross margin this quarter, PromptQL first showed its interpretation of the question, marking that it understood the acronym but had to assume what "this quarter" meant, then laid out a plain-language query plan.
Dominguez tagged a colleague into the thread to confirm when the fiscal quarter starts. PromptQL reran the query with the new information and suggested updates to its semantic layer, which a subject matter expert could review and save. The goal, he said, is a flywheel where the system signals uncertainty, people supply tribal knowledge, and each conversation improves the semantic layer.
See the next AI Demo Day
AI Demo Days are one of four series on Data Science Connect's 2027 schedule 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 sessions. Registration for 2027 events is coming soon.
Frequently asked questions
What is AI Demo Day?
A Data Science Connect virtual event where vendors each get about 10 minutes to demonstrate working AI products.
Why do AI agents need a checkpoint store?
Agents run multi-step processes, and saving state at each step lets a failed run retry from the previous checkpoint, as Aerospike's demo showed.
How can AI tools get location context?
Precisely demonstrated an MCP server that lets tools such as Claude Desktop complete and geocode addresses and retrieve property and nearby-place data.
How do you mask sensitive data in a real-time pipeline?
Striim showed agents that scan a source for sensitive fields and mask them in memory before the data reaches the target, such as BigQuery.
Watch: AI Demo Day Q1 on YouTube
Event: Webinars · More coverage: Webinars insights
