MAIKA Origin Story (P2): The Wilderness Years: Searching for the "Soul" in the Code
Date: 2017 - 2022
Phase: The Search for Connection
Mission: Teach the Robot to Be Human
We ended Part 1 with a soft victory: We had successfully automated the "grind." Our logic-based bots were answering thousands of questions and saving us hundreds of hours.
But there was a problem. Nobody likes talking to a robot.
We’ve all been there. You type a frustrated message to a customer support bot, and it replies with a cheerful, robotic: "I do not understand. Please select from the menu." It makes you want to throw your phone across the room.
We realized that "efficiency" without "empathy" is just a faster way to annoy your customers. So, in 2017, we started a new, ambitious phase: We wanted to cross the Uncanny Valley.
The Engineering Trenches: Wrestling with Intents, Entities, and Traits
For five long years, we didn't just "write code." We went to war with the complexities of human language.
In the era before ChatGPT, AI didn't magically understand context. We had to build understanding from scratch, piece by painful piece. We experimented with every state-of-the-art tool available at the time—from the accessibility of Wit.ai and the industrial-strength linguistics of spaCy, to the complex dialogue management of RASA.
Our engineering team spent thousands of hours manually stitching together the "DNA" of a conversation using the Holy Trinity of NLP (Natural Language Processing):
- Intents (The "Why"): We had to manually train the bot to distinguish between "I want to fly to Hanoi" (Booking Intent) and "Can I fly with a dog?" (Inquiry Intent). If the definition was slightly off, the bot failed.
- Entities (The "What"): We had to teach the system to hunt for specific details. Identifying that "tomorrow" is a date, "Danang" is a location, and "cheap" is a sorting preference. Tools like spaCy helped us parse these sentences, but it required endless tuning to handle Vietnamese slang and typos.
- Traits (The "How"): This was the hardest part. Using Wit.ai and later RASA, we tried to map the emotional temperature. Is the user sarcastic? Urgent? Angry?
It was a grueling process of "hand-crafting" intelligence. We were essentially trying to build a human soul out of math equations and decision trees. It was rigid, it was fragile, and it was hard work.
The Bottleneck: The "Message Generator"
Despite all this sophisticated engineering to understand the user (NLU), we hit a massive wall: Generation.
We could use RASA to perfectly understand that you wanted a refund, but we couldn't teach the computer to write a response that sounded... human. The "Message Generator" was our bottleneck. We were trying to teach a calculator to write poetry. No matter how good our logic was, the replies still felt stiff and "scripted."
The Pivot: If We Can't Speak Perfectly, We Must Listen Perfectly
Since the technology wasn't ready for fluent speaking, we decided to double down on our strength: Deep Listening.
We took all those years of struggling with Intents and Entities and turned it into our superpower. We realized that while we couldn't generate Shakespearean prose yet, we could understand the customer better than anyone else.
- We used our Entity models to catch details other bots missed.
- We used our Trait analysis to flag "Angry" customers instantly for human intervention.
- We learned to recognize the Buying Signal buried in a casual chat.
Building the Engine for MAIKA
Looking back, these "Wilderness Years" of wrestling with low-level NLP were the most important of our journey.
While the rest of the world was waiting for AI to get "smart," we were in the mud, building a massive library of Human Context. We were mastering the mechanics of conversation.
We had built the brain. We had built the heart. We were just waiting for the voice.
And in 2023, the voice finally arrived.
Next up in Part 3: The explosion of LLMs, the "Black Box" problem, and how MAIKA finally came to life.