Letting non-engineers train chatbots

Solo Front-End Engineer (UX-minded) Rasa
Rasa Platform: a conversations list, a chat transcript, and a panel for correcting a message’s intent

Rasa Platform is an ML-Ops tool for building and training enterprise chatbots, pre-ChatGPT. As the solo front-end engineer, I researched, designed, and shipped the React web app that cut handoff and training time by 40% and helped Rasa pivot from services to a product company.

Role Solo Front-End Engineer (UX-minded)
Timeline 2017–2018 · Berlin
Skills React web development · Sketch
  • 40% Less handoff and training time
  • Product-led The Berlin workshops moved Rasa from services to a product-led culture
  • Solo Researched, designed and shipped the React web app

The Problem

Enterprise customers face messy workflows to train and test ML models

While enterprise customers (like banking and insurance companies) had solid workflows to build chatbots in Rasa, training and testing models remained highly ambiguous.

Teams frequently hit roadblocks, relying on the Rasa team to build custom solutions so engineers and PMs could collaborate on iterations. Without these, engineers faced heavy overhead: parsing feedback from screenshots, emails and Jira notes.

To share a chatbot with a nontechnical teammate, engineers would even build custom sandbox environments, and sometimes, even teach PMs how to test chatbots using CLI terminals like a developer.

The Rasa Platform web app

How We Investigated

Learnings from existing customer relationships

In the spirit of Rasa making the shift from being a professional services company to a product company, I proposed to the CEO and COO that I run a design thinking workshop with the entire startup. I had everyone pull from their experiences building for existing clients.

We divided the existing clients into two personas: engineers and PMs. Each person mapped the behaviors, pain points, and goals of existing customers that build chatbots using Rasa.

Design workshop in Berlin Workshop participants mapping workflows

What We Learned

Five recurring pain points across all users

  • Make debugging training data easier?
  • Make annotating data less tedious?
  • Show how good the bot really is?
  • Smooth dev–PM collaboration?
  • Make deploying changes easier?

Early Iterations

Turning Pain Points into Design Opportunities

Solved the handoff bottleneck: developers couldn’t share work-in-progress models efficiently, and PMs couldn’t give actionable feedback. Created a shared submission flow with structured feedback collection that worked for both workflows.

Current flow Where Rasa fits in

The chatbot-training workflow across developer and project owner, shown before and after: the steps Rasa Platform takes over are highlighted in purple.
Low-fidelity wireframes for the platform
Low-fi wireframes, reviewed with engineering, business, and end users.
Early wireframes shared out for feedback

The Solution

A platform that lets PMs annotate and test independently

I designed the NLU training queue with confidence scores (so non-engineers knew what to fix), a chat + debug playground, and a share-via-link flow. Built it in React from wireframes to production under deadline.

Rasa Platform reviewing and correcting a bot's predictions
Rasa Platform: reviewing and correcting the bot's predictions, with confidence scores.
NLU training queue with confidence scores
The NLU training queue: confidence scores guide non-technical PMs to the data that needs fixing.
Stories Traces Confidence Scores NLU Training

Impact

Impact & outcomes

40% Less handoff & training time
Happy customers such as banking and insurance companies use it internally
Culture shift The Berlin workshops moved Rasa from services to a product-led culture

What I Learned

What I'd build next

Model performance visibility is crucial to the business. Next step: richer analytics, tracking quality by version, comparing confidence across intents. This work cemented my pivot from engineer to designer.

Other projects

© Parul Sharma 2026