AI Demo Day Q4: Mistral AI Apps, Snowflake Cortex Code and the matchmAIQer® Launch
Mistral AI showed how its new Apps feature takes vibe-coded tools to governed deployment, Snowflake's Cortex Code built, tuned and deployed a model with no hand-written code, and AI Connect launched matchmAIQer® for matching buyers with AI vendors.
By The AI Enterprise Magazine Team6 min read

Key takeaways
- Models need an app layer. Mistral's Studio Apps handles repositories, local testing, deployment, sign-in, logs and secrets, on Mistral's cloud, on premises or in a private cloud.
- Reuse beats rewriting. Mistral packages tested building blocks as capabilities; by Thomas's estimate the AI then writes about 11% of an app's code.
- Run ML where the data lives. Snowflake's Cortex Code profiled data, ran AutoML trials, registered the best model and deployed it as a REST endpoint, pausing for approval along the way.
- The agent's role is the user's role. Cortex Code assumes the user's privileges, asks before acting and runs in a sandbox without internet access.
- Vendor selection is a recurring job. In AI Connect's surveys, 83% of executives called finding the right AI vendors a significant challenge, the problem matchmAIQer® is built to address.
AI Connect's fourth-quarter AI Demo Day, a virtual event with headlining sponsors Mistral AI and Snowflake, gave each presenter about 10 minutes to show working software. Mistral AI demonstrated a new way to build and deploy internal apps on its Studio platform. Snowflake showed its coding agent building, tuning and deploying a machine learning model with no hand-written code. AI Connect closed the session with the launch of matchmAIQer®, its platform for matching enterprise buyers with AI vendors.
Presenters were Joffrey Thomas, Developer Relations Engineer at Mistral AI; Karan Sarao, Senior AI/ML Architect, Applied Field Engineering at Snowflake; and Brian Mink, Co-founder and President of AI Connect. Here's what each demonstrated.
Mistral AI: How do teams get from a model to a deployed app?
Thomas opened with Mistral's model lineup: Mistral Large 4, its newest and most capable general-purpose model, smaller general models, and specialist models for OCR, speech-to-text and text-to-speech, and content moderation, along with newer models for time-series forecasting and robotic navigation. But a model is only a means to an end, he said. Getting a product into production still means building the app, hosting it, deciding who can sign in and what they can do, and storing data and secrets safely.
Mistral's answer is Apps, a new feature in its Studio platform. Studio already handles API keys, team invitations, shared prompts, connectors, skills and secrets. With Apps, a handful of commands sets up a Git repository locally or on a remote GitLab, starts a local development environment for testing, and deploys the finished app. Each deployed app has its own logs and secrets, and sign-in is handled for you: users authenticate through Mistral and must belong to the same Studio workspace, though an owner can widen access to other teams or external users. Studio, and the apps built on it, can run on Mistral's cloud, on premises or in a customer's private cloud, which Thomas said matters for the enterprises Mistral mainly works with.
His examples were deliberately varied. One app tracks what people are saying about Mistral online. Another, built for a friend, searches job listings, scores them against a résumé and drafts a tailored résumé for each, written in LaTeX by Mistral Large and exported as a PDF. The slide deck he presented from was itself an app: a vibe-coded slides tool whose content he had asked Mistral's Vibe agent to build.
Enterprise work needs more than vibe coding, Thomas said. Mistral's forward-deployed engineers package reusable, tested building blocks, which Mistral calls capabilities, covering things like authentication, connectors, evaluations, guardrails, OCR and voice. They ship as TypeScript and Python packages with descriptions of what each does, and a skill helps an agent find the right ones for a project. By his estimate, the AI then writes roughly 11% of the code and capabilities supply the other 89%. A pre-recorded example assembled an analyzer for engineering diagrams from existing OCR and coordinate-extraction capabilities. When a need keeps coming up across projects, he said, that's a sign it should become a capability. Apps is available now.
Snowflake: What does agentic machine learning look like in practice?
Sarao focused on tabular machine learning, which he said still keeps the lights on in credit risk, fraud, churn and demand forecasting. Most of a modeler's time has gone to environments, dependencies, data preparation and feature engineering, with model fitting the quickest step. Agents change that, he argued, moving people from running workflows to steering them: deciding what matters, interpreting results and using domain expertise to check the agent's logic.
He named three things that hold teams back. Machine learning needs to run where the data lives, because moving data out to train adds latency and security gaps and loses governance. Development, inference and operations need to be connected end to end, so training knows about the feature store and the feature store connects to serving. And governance has to be built in rather than added later.
Snowflake's answer is Cortex Code, its coding agent, which runs in the Snowsight interface or from a command line. It can use models from OpenAI, Anthropic or Mistral, but Sarao said the harness, not the model, is what sets coding agents apart. Because Cortex Code runs inside Snowflake alongside the data, governance and compute, it doesn't need MCP connections to get context. Its machine learning skills cover the lifecycle, from distributed training and lineage to the feature store, batch and real-time inference and model monitoring.
In recorded demos, Cortex Code profiled a dataset in a notebook, then used its AutoML skill to build the best regression model for predicting customer spend over the next 90 days. It proposed a task type, metric and time budget and paused for approval. It then checked for leakage, class imbalance and outliers, ran six models at default settings (random forest performed best), tuned the top two and registered the winner in Snowflake's model registry. Sarao said the process typically takes 15 to 20 minutes with no code written by hand. Asked how to serve the model in real time, the agent drew an architecture diagram and proposed a deployment plan, including the compute pool, before creating a REST endpoint on Snowpark Container Services. A live run built a binary classification model from a prompt template stored as a Markdown file and fixed its own failing code along the way.
How does Cortex Code handle permissions?
Most of the audience questions were about control. Cortex Code always assumes the user's own role, Sarao said, so it can do anything that role allows and nothing more, and it asks permission before taking actions. It runs in a sandbox cut off from the internet, and its executions are logged. If an administrator removes a privilege mid-run, the next step fails and the agent reports the error so the user can decide what to do. For a question about independently verifiable evidence of what was authorized versus what ran, he said he would follow up with Snowflake's security details.
For teams with models already in production, he suggested starting with monitoring, notification policies, retraining pipelines and real-time scoring. Success is measured against the end state declared in the prompt, such as a target AUC, latency or queries per second. In regulated industries, users can tell the agent which features it may create, for example only attributes that comply with fair lending rules.
matchmAIQer®: How can buyers find the right AI vendors?
Mink introduced matchmAIQer® as AI Connect's first software platform. He cited AI Connect's surveys of more than 500 executives at its 2026 in-person events: 83% said finding the right AI vendors is a significant challenge, and 60% said they evaluate new AI tools or vendors several times a year or continuously.
A 15-question assessment of about four minutes produces an AIQ score, an AI maturity index benchmarked against the user's industry and all industries, and a list of matched providers from a catalog of more than 1,500. Users can also search in plain language. His example asked for a customer support agent for a retailer on Google Cloud and BigQuery that hands off to people in Zendesk and works in English and Spanish. Provider profiles summarize public discussion of each vendor, which Mink described as neutral rather than a recommendation. Users can compare up to four providers and request introductions.
The free tier includes the assessment, matches and a first report. Additional reports are $79 each, and the Pro plan, at $149 a month or $1,490 a year, adds 25-question deep dives for each focus area and unlimited reports. An Enterprise plan adds analyst-validated reports. Mink showed two report types: a landscape report for deciding where to invest across focus areas, and a report for a single area, such as data governance, for shortlisting vendors for an RFP. matchmAIQer® is available at matchmaiqer.ai.
See the next AI Demo Day
AI Demo Days are one of the series in AI Connect's 2027 lineup of free live webinars. See the webinar schedule for upcoming demo sessions.
Frequently asked questions
What is Mistral AI Studio Apps?
A feature in Mistral's Studio platform that sets up a repository, provides a local development environment and deploys apps with built-in sign-in, logs and secrets. Apps can run on Mistral's cloud, on premises or in a private cloud.
What is Snowflake Cortex Code?
Snowflake's coding agent. It runs in Snowsight or a command line, uses models from providers such as OpenAI, Anthropic and Mistral, and has skills for the machine learning lifecycle, from feature engineering and training to deployment and monitoring.
What permissions does Cortex Code have?
It assumes the user's own role, so it can do only what that role allows, and it asks permission before taking actions. It runs in a sandbox without internet access.
What is matchmAIQer®?
AI Connect's platform for matching enterprise buyers with AI solution and service providers. A 15-question assessment produces an AIQ maturity score and matched providers; reports and deep-dive assessments are available on paid plans.
This article is based on remarks by Joffrey Thomas, Developer Relations Engineer, Mistral AI; Karan Sarao, Senior AI/ML Architect, Applied Field Engineering, Snowflake; Brian Mink, Co-founder and President, AI Connect at AI Demo Day Q4 webinar on November 19, 2026. More coverage: Webinars.
Filed under: Agentic AI, Data & AI Infrastructure, AI Use Cases & Applications
Become a Sponsor







