Building Opera: Democratizing AI Model Creation
No-code custom AI model development
When ChatGPT launched, our users wanted to experiment with AI to simplify their tasks. I partnered with the AI Delivery team to build a platform for developing and testing simple models. The key challenge was helping users understand AI's limitations and possibilities. I worked closely with engineering, AI specialists, and eager beta testers throughout development.
Vision
Empower users to build AI models independently, reducing internal team dependency and accelerating innovation.
Problem to be solved:
Enable customers to build language models without technical expertise.
Business need:
Reduce reliance on the internal team and expand self-service capabilities.
My role:
End-to-end design leadership and adoption strategy.
Outcome:
Beta launch attracted two new customers specifically for this feature.
Learning:
Strategic hands-on guidance proved essential for successful adoption.
Opera is Cresta's no-code platform enabling contact center leaders to automate coaching and monitor compliance without AI expertise.
It allows users to create custom "if this, then that" rules for simple scenarios, complementing Cresta's AI models that handle complex interactions.
Prior to Generative AI, users relied on limited keyword-based triggers only.
Problem 1: Low accuracy
Keyword-only automation failed to accurately capture complex conversational moments.
Example: Detect when visitor is interested in new membership using keywords
Problem 2: Long wait
Users faced significant delays when requiring Cresta's AI team to build custom detection models.
Example: Detect when visitor is interested in new membership with a custom AI model
After extensive research and user interviews, we developed our first prototype and tested it with beta users. We simplified the process into three intuitive steps: 1️⃣ 2️⃣ 3️⃣ ... 🎩 🐰
First Iteration
In step 1 , user enters required information to train the model
In step 2, user marks what’s true or false to train the model
In step 3, user reviews the model predictions
What did we learn from the first iteration? 🤔
First iteration feedback 1
“I don’t speak this language”
First iteration feedback 2
“How do we ensure quality?”
Solution to feedback 1
You don’t have to speak the “AI language”
We eliminated AI terminology in favor of natural language. While this added some text, it significantly improved user understanding of the process.
Solution to feedback 2
Put on the ML engineer hat without knowing
1. Provide feedback on fewer but more meaningful examples
2. Ease into investigative mode
3. Empower users to make decisions at the right time
Final design
Less about speed, more about quality
Learnings
1. Clarity over speed
User comprehension proved more important than workflow efficiency for this AI toolset.
Guided introduction essential
Users described entering the AI space as "time-warping to a different dimension," making clear guardrails and instructions critical for adoption.