When we set out to build EnGenie, we werenβt just building a chatbot β we were crafting a conversational experience. One that felt less like talking to a machine and more like chatting with a helpful HR colleague.
Hereβs how we approached conversational design to make AI feel human.
Step 1: Understand The Userβs Intent
The first rule of good UX? Know your user.
We realized early on that not every query needed the same treatment. So we built a smart routing layer that could:
- Route policy questions to a RAG pipeline
- Send holiday queries to a SQL database
- Handle greetings with a lightweight LLM
- Manage follow-ups by stitching in past context
This intent-based routing made EnGenie feel smarter and more intuitiveβbecause it responded based on what the user meant, not just what they typed.
Step 2: Speak Like A Human, Not A Bot
Even when the answers were correct, something felt off. The tone was robotic, overly formal, or just⦠not helpful.
So we asked: What would an HR rep sound like?
Instead of fine-tuning the model (which is costly and rigid), we used prompt engineering to shape EnGenieβs voice. Our system prompt included:
- A friendly, personalized tone
- Direct address (βYou are entitled toβ¦β)
- Clear rules to avoid hallucinations
- A fallback message when info wasnβt available
This simple change made a huge difference. Suddenly, EnGenie didnβt just answer β it connected.
Step 3: Design For Context, Not Just Queries
Follow-up questions like βWhat about sick leave?β used to confuse the bot. Why? Because it lacked memory.
We fixed this by:
- Storing past messages in a conversation DB
- Merging old and new context before generating a response
Now, EnGenie could handle multi-turn conversations with consistency and clarity just like a real human would.
Step 4: Make The Invisible Visible
To improve UX, we needed to see what users were experiencing. So we integrated Langsmith for observability:
- Tracked query flows (RAG, SQL, etc.)
- Monitored token usage and latency
- Logged prompt experiments and outcomes
This helped us debug faster, optimize smarter, and ensure a smooth user experience.
Step 5: Measure What Matters
We didnβt want to rely on gut feel. So we used RAGAS to evaluate:
- Faithfulness (no hallucinations)
- Answer relevance
- Context relevance
- Context recall
This gave us a clear, data-backed view of how well EnGenie was performing and where to improve.
Final Thoughts
Building EnGenie showed us that a successful chatbot is more than just AIβitβs a human-centered experience. From understanding user intent to maintaining context across conversations, every design choice aimed to make interactions seamless and intuitive.
By integrating observability and measurement, we ensured that improvements are data-driven, not guesswork. EnFuse Solutions plays a pivotal role in guiding this journey, helping businesses create AI experiences that feel natural, reliable, and genuinely human.
