I’m deeply involved in building an AI product that translates complex data into natural, conversational language, leading strategy, UX, positioning, and team-building efforts while working directly with the engineering team. Here’s what I’m learning about the reality of AI development.

None of this works without the right people. We’ve assembled a remote team of A-players who can move fast, think strategically, and execute flawlessly. Having engineers who understand product vision and product people who respect technical constraints makes everything possible.

Designing How Teams Build AI Intelligence

As a designer, I leverage every tool in my toolbox. When we’re working on a challenge, it’s not just about identifying the exact problem and solving it; it’s also about understanding the underlying issues and addressing them effectively.

How we solve it becomes a means of transformation among team members, helping them see the dream and vision come to life. John Kotter said it best: “People change what they do less because they are given an analysis that shifts their thinking than because they are shown a truth that influences their feelings.”

The process itself increases buy-in and reduces lag because people experience the vision firsthand, rather than just hearing about it.

How we do THE thing is as important as what we’re doing and why we’re doing it.

Here’s what I’m most excited about: I’m designing how our team builds AI intelligence.

The Challenge: How do you let non-technical team members shape AI personality without requiring engineering skills?

The Solution: A simulation-to-production pipeline I’m orchestrating:

1. Simulate in ChatGPT: Product, design, and business teams build and test AI personalities using custom GPTs with dummy data
2. Extract the intelligence: Comprehensive prompts that reverse-engineer the personality into systematic instructions
3. Port to production: Engineering takes extracted prompts and deploys them via APIs

This approach is giving us:

• Rapid prototyping: Non-technical team members can iterate on personality in real-time
• Faster outcomes: No engineering bottleneck for personality development
• Quick PoCs: Prove concepts before heavy technical investment
• Cross-functional ownership: Everyone can contribute to the AI product’s personality, skills and functionality design

The Development vs Production Reality Gap

The Problem: Our AI works perfectly in ChatGPT. Rich personality, spot-on tone, brilliant insights. But when we move to production APIs, everything falls apart.

What I’m Learning: Development environments are forgiving. Production is ruthless. In the simulation environment (GPT), we have unlimited context and GPT-4o’s full power. In production, we hit token limits, model constraints, and cost pressures. The AI that feels human suddenly feels robotic.

What I’m Learning About AI Personality Architecture

Insight #1: Personality isn’t just “nice to have”
Personality is actually the product. Without it, you’re just another dashboard with a chat interface.

Insight #2: Token efficiency is everything
Our initial system prompt is 2,245 tokens. Our data payload is 61,000 tokens. The model can’t process both effectively. I’m solving this by breaking personality into tiers. Core DNA (300 tokens max) + contextual modules loaded as needed.

Insight #3: Model capabilities vary dramatically
GPT-4o vs GPT-4o mini isn’t just a cost difference. It’s a capability cliff. I’m learning to design for production models, not development fantasies.

The Data Intelligence Challenge I’m Solving

We started by dumping all store data into every conversation. Terrible idea.

I’m building a better approach:

• Query classification first
• Pull only relevant data subsets
• Progressive context building
• Intelligent degradation when needed

Most user questions need 5% of the available data, not 100%.

Working with Engineering: Strategy Meets Reality

As someone bridging product strategy and technical implementation, I’m learning:

Engineers see constraints. Strategists see possibilities.
The magic happens when you design within constraints rather than around them.

Process architecture matters as much as prompt architecture.
How your team builds is as important as what they build.

Production forces prioritization.
Every token matters. Every feature has a cost. You discover what’s truly essential.

The Extraction and Scaling Problem I’m Tackling

How do you capture an AI’s personality from a custom GPT and replicate it systematically in production?

I’m building comprehensive extraction frameworks to reverse-engineer personalities into production-ready prompts. This is teaching me that AI development is part psychology, part engineering, part linguistics.

You’re not just building software. You’re architecting intelligence and designing how teams create it.

What’s Surprising Me Most

The hardest part isn’t the AI. It’s the system around it.

• How do you maintain personality at scale?
• How do you handle edge cases gracefully?
• How do you balance sophistication with efficiency?
• How do you preserve brand voice under technical constraints?

These aren’t AI problems. They’re product design problems that happen to involve AI.

What I’m Learning About AI Product Development

1. Design your development process before you design your AI
2. Enable non-technical teams to shape AI directly
3. Build simulation-to-production pipelines for rapid iteration
4. Design for production constraints from day one
5. Token management is as important as feature development
6. AI products are systems, not just models

Building AI products means orchestrating how teams create intelligence, not just creating intelligence itself.

You need to architect both the AI and the process of building it. The companies that design effective development workflows will ship AI products faster and better than those that just focus on the models.

I’m also leading on strategy, positioning, design, and building our remote team of A-players, but that’s material for another post.

Note: Currently building conversational intelligence that makes complex data actually useful. Always interested in discussing AI product development challenges and solutions. Feel free to ping me if you’d like to chat.