Back to Portfolio

NeutonAI: From Ambiguous Idea to 92% Retention AI Platform

Building AI content tools that understand what people need without making them AI experts first

Here's the thing about content creation in 2025: everyone knows AI can help, but most tools feel like you're wrestling with a robot that doesn't quite get what you're trying to say. Businesses were spending hours tweaking prompts, managing multiple AI tools, and still ending up with generic content that needed heavy editing.

I saw this gap everywhere—in my work with outdoor brands, in conversations with marketing teams, in my own workflow. The question wasn't "Can AI create content?" It was "Can we build something that actually understands what people need without making them AI experts first?"

What We Built

The Product: An AI content-generation platform powered by RAG pipelines and vector search that cuts content creation time by 75%.

The Impact: 50 design partners, 92% weekly retention, estimated $1.2M in annual time savings for clients.

But here's what those numbers don't tell you—we built a system that turns ambiguous creative briefs into production-ready content without requiring users to understand transformer models or prompt engineering.

Starting With the Problem, Not the Technology

Most AI products fail because they start with "Look what GPT-5 can do!" Instead, I spent three months just listening. I talked to content marketers, agency owners, and freelance writers. The pattern was clear: they didn't need another chatbot. They needed a system that could:

  • Remember their brand voice across sessions
  • Pull relevant examples from their existing content
  • Deliver consistent quality without prompt gymnastics
  • Actually save time instead of creating a new learning curve

That insight shaped everything. We weren't building an AI wrapper—we were solving a workflow problem that AI happened to be really good at.

The Technical Architecture That Made It Work

Challenge: Getting RAG Right

Vector databases are great in theory, but getting relevant context retrieval right is hard. Too narrow, and you miss important brand nuances. Too broad, and you overwhelm the model with noise.

Our Solution: Built a hybrid retrieval system combining semantic search for conceptual matches, keyword search for specific terms and brand language, and recency weighting (because your brand voice from 2020 might not match 2025).

I worked directly with our ML engineers to test different chunking strategies, embedding models, and similarity thresholds. We ran 200+ experiments before finding the sweet spot.

The Framing That Changed Everything

Instead of asking engineers "Can you build RAG?", I asked "How do we make sure Sarah the marketing manager gets relevant examples every single time?" That framing changed everything.

The Beta That Taught Us Everything

We launched with 50 design partners—not because that's a magic number, but because it was small enough to have real conversations with each one while being large enough to spot patterns.

Week 1 Insight

People loved the output but kept asking "Can I save this as a template?" We hadn't built that feature because we thought the AI would make templates obsolete. Wrong. People wanted to capture what worked.

Week 3 Pivot

Added template saving, but made them "learning templates" that improved with each use. That single feature drove our retention from 78% to 92%.

Week 8 Discovery

Power users weren't using our suggested workflows—they were chaining outputs together in ways we never imagined. So we built a canvas feature that let them compose complex content from multiple AI passes.

This is the messy reality of AI product management. You can't A/B test your way to product-market fit when you're dealing with emergent behaviors. You need to watch, listen, and iterate fast.

The Contrarian Insight

Users don't want "AI"—they want magic that works.

Stop talking about your transformer architecture. Users care about: "Does this sound like my brand?" "Can I trust this not to make stuff up?" "Will this save me time or create more work?"

Frame everything around outcomes, not technology. The difference between a good AI product and a great one isn't the model—it's the product thinking around it.

The Numbers That Actually Mean Something

75%
Time Reduction
From 4 hours to 1 hour per blog post. Annual savings: ~$24K per user.
92%
Weekly Retention
Users came back every week because the tool actually saved them time.
$1.2M
Annual Savings
Estimated time savings across all 50 design partners.

Why 92% Retention Matters

This isn't about engagement theater. Users came back because the tool actually saved them time. Every week. That's the gold standard for B2B SaaS.

What I Learned About AI Product Management

AI Products Need Different Success Metrics

Forget your typical SaaS playbook. With AI products, you need to track:

  • Output quality scores (we used human evaluation + automated checks)
  • Iteration rate (how often do users regenerate? High = bad prompt interface)
  • Time to value (how long until they get usable output?)
  • Hallucination rate (tracked and improved from 8% to <2%)

The Importance of Human-in-the-Loop

Pure automation is a trap. The best AI products put humans in control while removing tedious work. We built review workflows, easy editing, and version control. Not because we didn't trust the AI, but because users needed to feel ownership.

Find 10x Value, Not 10% Improvement

Building NeutonAI taught me that AI product management isn't about understanding transformers (though that helps). It's about finding problems where AI creates 10x value, building the scaffolding around models (data pipelines, evaluation systems, user workflows), shipping fast and learning faster, and balancing autonomy and control.

Technical Stack

LLM: GPT-5 and Claude Sonnet 4.5 with custom fine-tuning for brand voice
Vector DB: Pinecone for semantic search
Orchestration: LangChain for prompt management
Infrastructure: AWS (SageMaker, Lambda, S3)
Monitoring: Custom dashboard tracking quality metrics, latency, cost per generation

The Bottom Line

NeutonAI isn't just another AI wrapper. It's proof that great AI products come from understanding customer problems deeply, building the right technical foundation, and iterating based on real usage.

We turned an ambiguous idea—"make AI content tools actually useful"—into a product that 50 companies use every week with 92% retention. That's not luck. That's product management.

Key Takeaways

Started with deep user research, not technology capabilities

Built hybrid RAG system after 200+ experiments to optimize retrieval accuracy

Achieved 92% weekly retention through rapid iteration based on user behavior

Focused on outcomes (brand consistency, time savings) over technical features

Want to Talk AI Product Strategy?

Let's discuss RAG implementations, beta strategies, or how to launch AI products that actually teach you something.

Let's Connect