Designing a Friendly, Accurate AI Chatbot

Reducing hallucinations, introducing name recognition, and upgrading to a smarter model.

Project Overview

I designed and developed a conversational AI chatbot to explore how UX design principles can make AI interactions more trustworthy, personal, and emotionally engaging.

Through several iterations, I improved accuracy, introduced dynamic name recognition, and upgraded to a stronger AI model (Mistral-7B).

Quick Stats:

🎯 Focus: Personalization + Accuracy

🛠️ Tools: Flask, Hugging Face Inference API, HTML, CSS, JS

🔧 Upgraded: BlenderBot → Mistral-7B

🕒 Timeline: 3 Weeks

What I Did

Designed and coded the chatbot interface using HTML, CSS, and JavaScript.

Built the backend using Flask and connected it to Hugging Face’s BlenderBot model.

Emphasized quick, clear user feedback — like loading animations and typing indicators.

The Problem

Early versions of the chatbot hallucinated answers and felt robotic.

❌ Incorrect or "made-up" AI responses

❌ Chatbot forgot the user's name

❌ Interface looked plain and unengaging

Major Iterations

Model Upgrade:
BlenderBot ➔ Mistral-7B

After testing BlenderBot, I realized the model was not instruction-tuned enough for friendly chats.
I upgraded to Mistral-7B-Instruct, using Hugging Face’s Inference API.

✅ Mistral followed prompts better ("Always call user by name")

✅ Responses became less random, more human-like

✅ Reduced hallucinations and improved trust

BlenderBot: Name Hallucinations

Mistral - 7B: Improved responses

Final Design Features

✅ Personalized greetings (remembers user's name)
✅ Controlled, polite AI responses

✅ Forgiving of typographical errors
✅ Typing indicator animations
✅ Mobile-style chat layout with avatars
✅ Secure environment with hidden API keys

Key Learnings

Visual cues (like a "bot is typing" animation) are crucial for keeping users engaged.

Error handling in chatbots needs to be graceful — users expect conversations to feel human even when they fail.

Small design details (like avatar icons, response timing) significantly impact user trust and comfort.