Developing an AI-based Recommendation Engine for India’s Largest Automobile Manufacturer

Despite the client’s vast user base of millions of users, the client was unable to share the entire user data with Team Daffodil due to data privacy protocols. Consequently, the client shared a limited data set with Team Daffodil. This posed a significant challenge for our team, potentially hindering the effectiveness of the recommendations.

To overcome this, Team Daffodil employed synthetic data utilization techniques & expanded the dataset leveraging machine learning (ML), to improve model performance. Our experts expanded the dataset, creating artificial data points that closely resembled real-world scenarios. This augmentation significantly improved model performance and accuracy, enabling us to deliver more effective recommendations.

On the basis of the targeted filters such as birthday, anniversary, maintenance due date, etc., the system showcases the list of highly probable/customers to whom the client could send the messages to.

The client could then individually view the customer’s profile, select the most appropriate details such as region, current car, budget, etc. and send customized messages from the list of available, AI-generated options, without having to scrounge through the vast list of users manually or type individual messages every single time for each user.

The recommendation message was either a new car suggestion for users who’ve owned the vehicle for over 5 years; or a maintenance-related message for recent owners.

Car Recommendation Engine:

Leveraging collaborative filtering, the car recommendation engine offers personalized vehicle suggestions to customers, based on their past purchase history and preferences.

In order to predict if a customer would like a product recommendation or not, the system employs matrix factorization and forecasts how the customers would rate the cars even when they haven’t interacted with them physically. It analyzes data types such as user data (age, gender, location, etc.), car-related information such as pricing, fuel efficiency, etc., past purchases and models owned, along with feedback on test drives, if they’ve taken any.

Further, the recommendation engine is trained with user profiles to offer relevant suggestions.

True Value Prediction:
The solution also offers the feature to predict the true value of a customer’s current vehicle in case they’re looking out for an exchange/upgrade or a trade-in opportunity. Leveraging the same, the system can offer relevant messaging to customers and help to increase customer retention and loyalty towards the brand.

Best Time Prediction:
To enhance the efficiency of responses, Daffodil leveraged the Generative AI capabilities of Large Language Models (LLMs) along with automation to generate human-like messages. Through a rule-based system, it became possible to predict the best time to reach out to customers.

Here’s how the rule-based system works:

Special Occasions: The rule-based system identifies special days such as car purchase anniversaries, birthdays, etc. to send messages around that time for maximum user engagement.

Festive Days: The system schedules messages and greetings around major cultural holidays to connect with users & resonate with their religious practices.

Maintenance Reminders: By analyzing a customer’s vehicle purchase date and maintenance history, the system predicts the right time to send service and maintenance reminders to the target customers.

New Offers: For customers who are due for an upgrade or have shown interest in newer car models, the system directs regular offers, launch info and promotional messages to those users.

Message Generation:

The LLM-based message generation process collates the data and insights and transforms them into personalized messages leveraging GPT-3.5. Team Daffodil had set clear guidelines while training the solution to showcase results based on the brand’s style of language and emotional undertones such as professional, friendly, exciting, etc. This was done to generate brand-like tonality in the messages.

Additionally, by leveraging prompt engineering, our team enhanced the solution by utilizing content-based filtering & eliminating certain keywords that did not adhere to the client’s brand policies. These phrases and keywords were then integrated into the LLMs’ operational framework.

The recommendation engine was then trained with previously successful messages which were further personalized for the customer’s unique needs.

For the solution to work seamlessly, the system is made to intelligently integrate specific data such as car recommendations, service due dates, etc. at a suitable time in a natural, human-like tonality that feels personal to each user.