Creating digital employees with the help of open-source LLMs.

1. 1.

   Since this is a service business operating on a home-service model, it is essential to define the service coverage area; based on current operations, this is determined at the sub-district level.

2. 2.

   Since this is a service business with pre-set prices, I need a system capable of interpreting or normalizing customer language.

3. 3.

   In addition to customer language normalization, I also need a system capable of reading and interpreting images or content sent by customers, in order to identify and match them against the price list. Beyond price list issues, I also require a system that can interpret images or videos based on customer complaints.

4. 4.

   I also need data to conduct research for business development and the development of this system itself.

5. 5.

   I need a WhatsApp API that can be integrated with the system I am building.

6. 6.

   I need a database to store all customer data, price lists, complaint types, service areas, and schedule bookings.

7. 7.

   I need a system capable of reading specific customer house numbers to determine whether or not they are detected.

8. 8.

   I need a payment gateway that can be integrated with my system so that every payment can be verified in real-time.

I set out to build this system by leveraging AI to the fullest, using a sufficiently large LLM to bridge the gap regarding technical syntax requirements and integrating AI into every possible node to keep the system streamlined and minimize the node count. However, my reasoning turned out to be completely flawed, and the hundreds of tests I conducted using this approach failed utterly.

1. 1.

   The AI I installed hallucinates a lot 🤣. Its AI often fails at crucial or critical points, such as cost estimation, coverage validation, and so on.

2. 2.

   The system has become unpredictable; it is difficult for me to determine which parts actually require mitigation, which are erroneous, and which are broken — it is simply too complex.

3. 3.

   The payment gateway — and this is the most crucial point — its AI consistently fails to identify whether the payment has actually been made or not.

Reflecting on the failures I've experienced, I've come to realize that the answer isn't always simply "using a more capable AI to build the system and shorten the journey." Those setbacks led me to shift my approach from "AI-first" to "deterministic-first," reserving AI only for instances where it is truly the right fit.

The results I achieved are the product of hundreds of tests and dozens of iterations and methodological adjustments. The outcome is a system that is easy to debug and mitigate; customer inputs not detected by the system are captured in the database to inform upgrades, anomalous inputs are logged for market research, and so on.

AI is powerful, but its true value lies in using it at the right time, leveraging its strengths, and most importantly knowing how to use it effectively. Through my own experiences with failure, I've learned that AI can be a highly effective tool for running a business like the one I'm currently building without requiring massive resources, provided we understand how to use it and identify the right approaches.

This is not a simple task, nor is it something a beginner can do "easily." If you want to build something like this, start by creating a system where the AI parses keywords and generates simple reply templates — from there, you will likely feel challenged and want to take it further.