The customer wanted to improve demand forecasting accuracy by implementing AI/ML based approach. Currently the forecasting is done in Excel with gut feeling than following any scientific approach. The current forecasting accuracy is around 60-70%
Python, Flask, Time Series Algorithms
The customer is a leader in water heater category in India. We have received product sales data at day level granularity for last 2 years. The data is tested to check missing values and outliers. The products were categorized based on the historical monthly sales data. Using feature engineering, the missing values were treated first. Using Dickey-Fuller test, the data was tested for Non Stationarity. The data did have trends, and seasonality, which was treated and the data was made stationary.
Later, various time series algorithms like ARIMA, Holt-Winters, XgBoost &LSTM were applied to the data., Holt-Winters gave an accuracy of over 90%, which was fine tuned by tuning hyper parameters. Other algorithms gave accuracy in the range of 75% to 85%
The customer’s customer ( An Organization of Govt. of India), wanted to do a Preventive Maintenance Proof of Concept for Air Conditioners. They asked to prepare a report on a set of data generated using various test applied on the Air Conditioners.
R, Data Visualization, K Means Clustering
The supplied data was provided only for 18 hours with details like Voltage, Current, Vibration, Compressor Temperature, Room Temperature, and date time component.
The data did not have a clearly defined breakdown condition. Neither the data was qualified for time series forecasting. It was clearly a case of un-supervised learning. We also for relationships between the variables and found that only Current, Vibration and Compressor temperature are good enough to represent the data set. We then applied K-Means Clustering algorithm to the data to create segments with data points which have more or less similar characteristics. With a K-Value of 15, we could distinctly identify on cluster which more or less, had data points, which were defining probable Air Conditioner breakdown condition.
A report on the findings was created with 3D visualizations of the Clusters along with our findings and submitted to the customer. We requested for more clearly defined breakdown conditions which can be used for more accurate predictions.
The customer is building a platform for automated buyer seller negotiation. The current version of the platform is relying on rule based engine. The customer wanted to use AI auto price discovery during buyer seller negotiation.
Python, Scikit-Learn, Flask, Tensorflow
Worked closely with the start up in North America to define how AI can be introduced in their current platform. Provided use cases relevant to the business domain and helped to create roadmap for AI implementation
Developed the model for buyer classification and seller classification to get intelligence about buyer & seller. Developed descriptive analytics solution for getting intelligence about previous deals to understand whether they were buyer centric or seller centric. Created API’s in Python flask to expose AI models as a rest web service. The next phase of the solution involved automating price discovery and buyer/seller negotiation.
The end client, a Solar Power Plant, had to submit the power generation forecast for next 7 days to Govt. of Karnataka, as a part of compliance reporting. The client did not have any in-house capabilities in machine learning based forecasting.
Python, Time Series Forecasting Models
The client had Scada system implemented for Solar Power plant automation. The solar power generation data was getting collected for each of the solar panel. The data collected was then sent to cloud at an interval of 15 minutes. The client had the historically collected data for over 24 months. The data was in the form of time series.
The scope of the work was to training the machine learning model on the historical data for the model to understand the underlying data patterns. The historical data, along with the weather data was used to forecast the power generating for next 7 days, and push the data to the designation cloud server.
The ML models tried were ARIMA & LSTM. The model accuracy was tested by comparing the forecast with the actuals for a period of 40 days. We could get over 85% forecasting accuracy using LSTM model. The mandate was for 80% forecasting accuracy, thus we exceeded the clients expectations.
The client was looking for improving recommendation and search experience based on user behaviour. The client wanted to reduce dropout rate and engage customer in better way to drive customer loyalty
Python, Scikit-Learn, Flask, Tensorflow, SQL Server,ElasticSearch
Implemented multiple strategies to recommend right product to the customer based on previous buying behaviour and buying pattern of customer with similar profile. Enriched customer data with additional attributes by understanding buying patterns. The customers are micro segmented using customer segmentation model. This is used for understanding customers which are loyal customers and which may churn. All this information is used to define better product promotion strategy.
As the data was less when we started the engagement, the synthetic data was generated with the help of domain experts. This has helped to significantly improve accuracy of AI models. After implementation of the AI models, we observed significant uplift in revenue and dropout rate decreased significantly.
The grocery ecommerce start up was looking to improve search experience of the customer. The current search engine is keyword and tag based. It does not take into account semantic relationship and it was not personalized based on customer behaviour and preferences
Python, Machine Learning, Recommendation Engine, BERT, ElasticSearch
The solution used ElasticSearch for implementing the search functionality. Instead of keyword based search, semantic search functionality is used where word embeddings are created using BERT and used vector representation in embedding for the search. The pipeline is developed to understand customer behaviour and developed Machine Learning based solution to contextualize the search & make it more personalized. The implementation of new search engine helped to greatly reduce drop out rate of the customers from the site and improve conversion ratio