During Abnormal’s AI Ascent APAC hackathon, Tushar Amrit built a smarter feedback loop for Nora, our internal AI assistant. In just two days, he turned simple emoji reactions into a structured system that helps Nora learn from every conversation.

Closing the feedback gap

Before this project, Nora’s feedback relied on Slack emoji reactions. They were easy to click but hard to learn from. Teams couldn’t see how Nora performed or whether her answers helped users. The data sat unused, leaving support channel owners guessing.

Three main frictions emerged:

1. Emoji-only feedback lacked context or reasoning.

2. No analysis loop to improve Nora’s prompts.

3. Missing or outdated documentation caused Nora to hallucinate.
   

Without a way to act on user signals, Nora couldn’t improve meaningfully over time.

Tushar’s AI tool added two major upgrades to Nora’s ecosystem: rich feedback capture and automated weekly performance analysis. As he explained, “This project was about AI improving AI.”

Core capabilities include:

• Interactive feedback pop-up that captures reasons behind a rating and optional comments.

• Automatic sentiment extraction from user replies in threads.

• Weekly channel reports showing message accuracy, user sentiment, and when human help was needed.

• System prompt recommendations suggesting updates to improve each channel’s Nora persona.

• Documentation gap detection highlighting missing or outdated internal references.

Together, these features create a closed learning loop. Nora doesn’t just record reactions; it interprets them, identifies weak spots, and recommends targeted improvements.

This means each week, Nora and her human owners both get smarter.

With the new analysis layer, teams can now see exactly how Nora performed. Reports summarize every conversation: how many were correct, needed confirmation, or went off-track. Sentiment data drives star ratings for each interaction, while leaderboard insights motivate users to provide better feedback.

So far, results are clear:

• Real-time accuracy tracking at the thread level.

• Reduced hallucinations by exposing documentation gaps.

• Improved response quality through auto-suggested prompt tweaks.

• Higher team engagement via weekly feedback leaderboards.
  

These insights let AI pilots like Albert quickly identify improvement areas across their channels. The system blends human intuition with machine learning feedback loops, making Nora incrementally better without manual tuning.

Next, Tushar plans to expand this loop across more internal channels to benchmark performance across personas.

Early users describe the experience as “AI helping AI.” Instead of static updates, Nora continuously self-corrects based on live user data. Teams now trust that feedback directly shapes future performance, creating a culture of shared ownership over AI quality.

This project captured Abnormal’s builder mindset of practical innovation that compounds. By connecting user feedback, analytics, and documentation hygiene, Tushar’s solution ensures Nora doesn’t just serve the team; it learns from them. The next step is scaling these loops so every persona can evolve with the same intelligence.