Foundations, Not Features: Inside the First Marketverse Vibe Coding Workshop

Why This Workshop Existed

Beau opened the session by naming the problem most attendees already lived with. “I’ve been going to a lot of these marketing get-togethers recently,” he said, “and one of the topics is always, what are you doing to answer the pressure of the board being like, use AI? And everyone’s kind of like, I can do something, but it’s not connected to anything. It doesn’t work with any of my systems.”

The Marketverse Vibe Coding Workshop was conceived as a monthly antidote: a small, intentional community where marketing leaders could see foundations built live, warts and all. No polished demos. No pre-baked outcomes. Just real construction in real time, with real errors to debug along the way.

What were they actually building?

The chosen scope was a proof of concept that touched every common request from the group: input a target list of accounts, pull in their tech stack and funding data, fetch earnings call transcripts, run an AI analysis to score fit, publish the result to a micro-site, and push real-time alerts to Slack. By the end of the hour, the team had stood up roughly sixty percent of that scope, including the company input form, three working API integrations, a Supabase backend, and a Claude-powered fit-scoring analysis.

Signal Integrity, And The Alpha Signals That Set You Apart

Before the build began, Lara grounded the room in a framework she returns to constantly: signal integrity. She has watched marketing teams cycle through CRMs, enrichment tools, and orchestration layers, only to keep landing in the same place. Signals everywhere. Action almost nowhere.

Her three pillars define whether a signal is trustworthy enough to act on:

From signal integrity, Lara moved to what she called alpha signals, the unique business intelligence that sets one organization apart from the rest of the market. Her example was a financial services company selling to registered independent advisors with five billion dollars or less in assets under management. Everyone has the same generic alerts now, she explained, job changes, earnings statement mentions, intent feeds. The top sellers at that company had learned something else: when the third call included a Chief Investment Officer joining the advisor team, the deal almost always closed within a week.

That kind of pattern is the alpha signal. It belongs to the people closest to the work, and it is the layer where AI can finally do useful orchestration once you know what you are looking for.

Lara was clear about why this matters now. The market has democratized the generic signals. Everyone sees the same job change alerts, the same funding round notifications, the same earnings call mentions. The competitive edge has moved underneath that layer. It lives with the seller who notices a specific team composition. The CSM who recognizes a deployment pattern that always precedes expansion. The marketer who knows which content asset a buyer reads right before they convert. Those signals are alpha because they belong to your business and almost nobody else has them yet.

The Build, Step By Step

With the framework set, Beau opened a blank project and began the live build. He chose Bolt as his vibe coding tool, with Supabase as the database underneath. The reasoning was practical rather than tribal. Lovable, which several attendees including Hannah preferred, keeps the database inside its own environment. Bolt lets the database live in Supabase directly, which means edge functions are visible, error logs are inspectable, and the whole stack can be claimed and moved if needed.

Why Supabase instead of a traditional server?

“Edge functions are basically, I have my Excel table or whatever, how do I get that data out and in from other places,” Beau explained. “Used to be you would have to set up a server and that server would be the communication between these two things. But thank god we live in 2026 and now these services have built-in features where you can interact with the data with the APIs.”

The first integration was Alpha Vantage, the free source the team picked to pull earnings call transcripts. Beau pasted the API documentation into the LLM, asked for an edge function that would call the endpoint and log any errors back to the UI, and watched it generate. The function dropped into Supabase. He plugged in the API key as a secret. A button appeared on the front end. He clicked it. The transcript loaded.

The first thing he flagged for the room: error logging. “When you’re using AI to generate code, you have to specifically tell it to log the errors. Otherwise it’ll just say unknown error, or it’ll just say error 500. And then you have to dig around in a hundred other places. It’s just easier to have it right in your face.”

Then came LeadMagic, with two more endpoints in a single pass.

Next came LeadMagic for funding rounds and tech stack data. Beau pasted both endpoint examples into the prompt and asked the LLM to wire them up alongside the earnings call view, this time with a Claude API key included so the system could analyze the consolidated information. The build took longer to generate but landed cleanly. When the page rendered, Snowflake’s profile correctly surfaced Databricks as a competitor.

The JSON Discipline

At the midpoint of the build, Beau paused to make a point he treats as non-negotiable. “If you send unstructured information to an LLM, it’s not every time, but it’s likely enough to hallucinate that you’re going to have an issue.”

The discipline is to store every external response in a structured format, JSON in his case, so the LLM is not interpreting prose. When the model needs to know whether Snowflake is public, it does not infer from a paragraph. It looks at the field. That structure is what makes the system scale predictably across hundreds or thousands of companies without drifting.

Lara picked up the thread here, connecting the JSON discipline back to her framework. Structure is what gives lineage and semantic consistency something to live in. Without it, every model run is starting from scratch, interpreting prose, guessing at definitions. With it, the system can hold the same meaning across time, across runs, across teams. That is the difference between a tool that hallucinates confidently and one your sellers will actually trust enough to act on.

A question on API key hygiene from Megan Cabrera.

Megan asked how Beau managed keys across multiple projects. His answer was deliberately low-tech: he does not store them anywhere long-term. He generates a key, drops it into the Supabase secret, and lets it live there encrypted, where no one, including him, can read it back. If a problem comes up, he regenerates. For teams managing dozens of keys across mature applications, he mentioned Doppler as a real solution. For the scope most marketing-led builds operate in, regenerate-when-needed is the cleaner habit.

Picking The Product: Cortex For Snowflake

With the foundation in place, the group needed a fictional product to anchor the AI analysis. Ruthie suggested a usage optimization tool for Snowflake clients, modeled on the kind of cost-and-pricing transparency her own team provides for Databricks customers. The room agreed. The product was named Cortex, and Beau handed Claude a short brief: generate five plausible services, then score each target company on fit against them.

The output came back surprisingly coherent. API cost intelligence. Token budget governance. Semantic caching. Data product metering. LLM routing and model arbitrage. All five could plausibly exist as a real product line for a Snowflake-adjacent optimization company. “Maybe maybe we should start this company,” Beau said. “We’ll be able to sell the absolute hell out of it, but it doesn’t exist.”

Lara noted what was actually happening underneath the surface. The fit score was not magic. It was the result of a structured comparison between two well-defined inputs: a service portfolio with clear value props, and a company profile with funding, tech stack, and earnings context. Both were stored as JSON, both were time-stamped, both could be traced back to their source. Semantic layer, temporal, lineage. The same three pillars she had described an hour earlier were now operating quietly inside a working app.

Temperature, And Why The NYC AI Officer Sets It To Zero

One of the smaller but more practical moments came near the end, when Beau explained the temperature setting most marketers never touch. Higher temperature equals more creative output. Lower temperature equals more deterministic output. For an analytical task like comparing a company’s funding data against a service portfolio, creativity is the enemy.

He shared an anecdote from a recent panel with the Chief AI Officer of New York City, whose team analyzes city codes and laws. “He’s like, dude, I’m in the government. Zero. There’s no creativity involved in this.”

Beyond The Build: AEO, Visibility, And The ICP Layer

Hannah Siegertsz steered the closing discussion toward Answer Engine Optimization. How do people find your product when the discovery layer is no longer Google but Claude, ChatGPT, and Perplexity?

Lara, who has been building a reporting tool for clients on exactly this question, pushed back on the framing. She prefers the term AI visibility, because the work goes deeper than updating robots.txt and structured data. Her reports generate roughly twenty pages of audit and six pages of specific actions, things like missing Wikipedia entries, robots.txt configurations, content gaps the answer engines cannot index. Some of those actions are pure automation. Others require humans deeply involved in content. The goal is to set up the system so the orchestration happens daily, because the underlying landscape changes every week.

Beau added a technical note for anyone building customer-facing applications: pay attention to whether the framework renders on the client or the server. React is client-side rendering, and search engines, including the answer engines, still struggle to access content that only assembles in the browser. Next.js, which renders on the server, is the safer default for anything that needs to be discoverable.

Mark Barrera closed the loop on strategy. Whatever the technical implementation, AEO begins with knowing exactly who you serve and what they actually ask. “I do a lot of call transcript mining, customer service, sales, whatever you may have to figure out who the people are, what they’re asking. Because that’s what they’re going to go do in these prompts.”

The technical layer is necessary. The ICP layer is what makes the technical work matter.

When And Where The Series Goes Next

The Marketverse Vibe Coding Workshop meets on the third Wednesday of every month. The next session is scheduled for Wednesday, June 17, and will build HubSpot integration on top of the Cortex foundation from this one, including auto-generated contact records, scoped API permissions, and pipeline routing from the AI analysis layer.

What Actually Came Out Of The Hour

Stripped of the live-build drama, the workshop delivered something specific. Not a finished product. Not a polished application. What it delivered was a working illustration of a foundation pattern: a structured database holding company records, three external API integrations feeding that database, a Claude analysis layer reading the structured data and writing structured output back, and a front end thin enough that it can be replaced or extended without touching the underlying logic. Snowflake was just the example company the group chose to run through it.

“This is the foundation,” Beau said as the hour wrapped. “If you can get to this point, this is when you can build basically any of your alpha signal intelligence, analysis, and push it from anywhere. Everything that I have built so far is based on this exact foundation. Just the information I’m inputting has changed and the way I’m outputting it. But the structure of it is exactly the same.”

The broader point Lara and Beau kept returning to is harder to put in a slide. The internet, Beau argued at the end, has become difficult to navigate because so much of what is published is AI reading AI, with hallucinations compounding hallucinations. Small, intentional communities of practitioners who actually build and share what they build are becoming the genuine resource. That is what the Vibe Coding Workshop is being built to be.