From Generic GPT to Proprietary AI Asset: An AI Persona Tool Case Study

A three-month build that turned an agency's tactical workflow problem into a productized, proprietary AI asset.

The Situation

A tech-forward creative agency had been using off-the-shelf custom GPTs to simulate customer focus groups for one of their largest enterprise clients—a Fortune 500 automotive brand running a high-traffic peer-to-peer marketplace.

On paper, this was a smart workflow. Run a concept past a synthetic audience before spending real research budget to test it with real customers. In practice, it kept breaking down on the same two failure modes.

The first was sycophancy. Generic LLMs are trained to be helpful and agreeable, which means they tend to praise whatever idea is put in front of them. For an agency trying to stress-test creative concepts, that's the opposite of what's useful. You don't need an AI that tells you your work is good. You need one that tells you where it's weak.

The second was missing audience nuance. Public LLMs had no access to the agency's first-party research, no access to their licensed audience insight reports, and no memory of historical content performance. The "personas" they simulated were stereotypes pulled from training data, not the actual modeled audiences the agency had spent years understanding.

The agency came to FM with a clear ask: build a proprietary AI persona tool that defeats sycophancy, grounds itself in real research, and—critically—is an asset they own rather than a third-party product they rent. They wanted to walk away from the engagement with software they could extend, deploy across their book of business, and use as a strategic differentiator against other agencies.

The Approach

FM staffed the engagement with a small senior team and ran it as a three-month build. The technology stack was deliberately chosen for AI-native workloads and long-term maintainability: Next.js and React on the frontend, the Vercel AI SDK for LLM orchestration against OpenAI's frontier models, Neon Postgres with pgvector for the retrieval layer, PostHog for analytics, and Vercel for hosting.

But the technology was the easy part. The hard part—and the part the agency was actually paying for—was the AI engineering.

Defeating Sycophancy by Design

Sycophancy isn't a bug you fix with a single clever instruction. It's a default behavior baked into how LLMs are trained, and it has to be engineered around at multiple layers. So FM didn't try to solve it with one prompt—the persona system was built as a layered architecture, where each layer pushes the model further away from generic agreeability:

A "Research Lead" system prompt layer that frames every interaction as a structured research exercise rather than a creative brainstorm, orienting the model toward critical evaluation rather than encouragement
A "Master Tone of Voice" layer that defines how every persona communicates—skeptical, direct, grounded in their own perspective—regardless of who's behind the keyboard
"Anti-Sycophancy" guardrails that explicitly prevent the model from parroting the user's input back at them or copying example content verbatim
A structured persona model (covered below) that gives the AI a real point of view to reason from, rather than defaulting to an agreeable generic baseline

The result is a tool that disagrees, pushes back, and surfaces weakness—the things a strategist actually needs from a "first filter" before live customer testing.

Personas That Reason from Mindset, Not Stereotype

The second core problem was the persona model itself. Most "AI persona" tools are little more than demographic templates with a prompt. FM designed something deeper.

Every persona in the tool is built on a four-pillar framework: Demographics, Psychographics, Behavioral Data, and Response Guidelines. And critically, the weighting between those pillars is calibrated to reflect how real audiences actually behave.

Psychographics carry 40% of the weight. Behavioral data carries significant weight. Demographics carry just 10%.

That weighting matters. It forces the model to reason from mindset—motivations, attitudes, decision-making patterns—rather than from age and zip code. The flagship persona built on this framework was a high-fidelity model of 25- to 41-year-old private vehicle sellers, the exact audience the agency's enterprise client needed to reach with their peer-to-peer marketplace.

Grounded in Real Research, Not Model Imagination

A persona is only as good as the data behind it. So FM built a Retrieval-Augmented Generation (RAG) pipeline on Neon Postgres with pgvector that ingests the agency's actual research documents and licensed audience insight reports.

When a strategist asks a persona to react to a creative concept, the system retrieves the relevant research, grounds the persona's response in actual data, and—this part matters—shows the strategist exactly which documents the AI referenced.

Source transparency is what makes the tool trustworthy for high-stakes strategic decisions. It's the difference between "the AI said this" and "the AI said this, and here's the research it pulled from." For an agency selling strategic rigor to enterprise clients, that distinction is non-negotiable.

A UI Built Around How Strategists Actually Work

The first version of the persona builder was a rigid, form-driven data entry experience. It was clean and well-structured, and it didn't match how the agency's strategists actually thought about persona construction.

FM watched the team work, threw out the form, and rebuilt the experience around a three-tab structure: Demographics & Psychographics, System Prompts, and Tone of Voice. The new UI mirrored the strategists' actual mental model—building a persona is a craft, not a data entry task—and adoption inside the agency followed immediately.

The full application includes:

Interactive chat for real-time evaluation of text and image-based creative
A searchable knowledge base of uploaded research documents
An admin dashboard with persona management, document tagging, and full version history with rollback
Source transparency built into every persona response

Multi-Tenant from Day One

The agency's ambition wasn't just to solve the problem for one client. They wanted a platform they could deploy across their broader portfolio.

FM designed the database, authentication, and access model to be multi-tenant from the first commit, with three tiers of access (User, Admin, Super Admin) and clean separation between clients. The Fortune 500 automotive brand was the first deployment. Additional enterprise clients have been onboarded onto the same infrastructure since.

That decision—designing the platform as a product rather than a one-off—is why the agency now has an asset they can sell against, not just a tool that solves one workflow.

Compound Engineering for Velocity

A three-month build for a proprietary AI persona platform with a custom RAG pipeline, multi-tenant infrastructure, and a full admin experience is aggressive by any measure.

FM hit the timeline by leaning hard on compound engineering—automated workflows powered by Claude and MCP that collapse 4+ hours of manual strategy and prompt-engineering work per cycle into parallelized AI sessions. Human-in-the-loop direction on architecture, AI velocity on implementation. Same model FM uses on every senior-team build.

The Result

The platform shipped on schedule, with final client approval, and zero post-handoff revisions requested.

What got delivered:

A proprietary AI persona platform built on Next.js, Vercel AI SDK, Neon Postgres + pgvector, and PostHog
A high-fidelity flagship persona modeling the agency's enterprise client's most strategically important audience segment
A multi-tenant infrastructure that now serves multiple enterprise clients, with more being onboarded
A complete infrastructure handoff: Vercel hosting, GitHub repositories, and PostHog analytics fully migrated from FM's environment to the agency's internal stack
Technical documentation for the agency's IT team to maintain and extend the platform long-term

Strategists can now iterate on creative concepts in minutes rather than waiting for traditional research cycles. The tool functions as a "first filter"—catching weak concepts before they consume live research budget and putting the strongest work in front of real customers.

And the agency walked away owning the asset, not renting it.

Why It Worked

Layered engineering against the core LLM weakness. Most "AI persona" tools are a single prompt and a clever UI. FM engineered the persona system at multiple layers—system prompts, tone-of-voice, anti-sycophancy guardrails, and a structured persona model—because sycophancy can't be solved with a single instruction.

RAG grounding with source transparency. Trust is what makes AI output usable for strategic decisions. Showing the receipts is how you earn it.

Calibrated weighting that prioritized mindset over demographics. Personas reasoned from psychographics and behavior, not from stereotype.

A UI built around the strategist's workflow. The form-driven first cut got thrown out. The three-tab structure matched how the team actually worked—and adoption followed.

Multi-tenant from day one. The agency didn't pay for a one-off. They got a platform they could productize.

A clean infrastructure handoff. The agency owns the code, the database, the analytics, and the deployment pipeline. No vendor lock-in to FM. That's the model.

For Agencies Considering Their Own AI Platform

If you're an agency still renting capability from third-party AI tools, this is what's possible when you invest in building your own.

You don't need a fully staffed in-house AI team. You don't need a year-long roadmap. You need a senior team that understands both the engineering and the strategic problem you're trying to solve, and a willingness to build something defensible rather than buying something generic.

That's what FM builds for.

Ready to Build Your Own?

If you're ready to stop renting AI capability and start owning a proprietary AI asset, FM can help you ship a production-grade platform on a timeline you'll actually hit.

Get Started or Learn More About Our Work