Lalu Prasad Lenka

I performed profit erosion analysis using multivariate descriptive and regression analysis to find the root cause of profit erosion across five major sub-brands of a confectionery giant.

Afterward, I discovered loss-making consumer pack types and adverse cost components. The next task I did was to perform a prescriptive analysis, and then I suggested business actions that optimized existing revenue streams by 18%.

Soon after, I performed prescriptive analytics and used time series forecasting to project the sales, major cost components and predicted the potential losses if no action taken. Finally, I created a "What if Tool" to prescribe the next best business action.

I developed and deployed an intelligent supply chain solution for a fast-fashion client that recommends more favorable SKUs and forecasts their 14-week sales and required inventory.

The solution helped the client maintain optimal stock levels for all SKUs and increased profitability by 15%.

I developed and deployed an LSTM-based named entity recognition model to detect complex chemical names and medicines in drug formulation documents that helped reduce manual efforts by 90%.

The model was developed on the IUPAC dataset and achieved an F1 score of 0.85.

I developed and deployed a deep learning pipeline to locate and recognize texts in machine screens like temperature and pressure and log in directly to the database, which helped reduce human error and operations costs by 40%.

Text extraction on an image was done in two independent steps—detection (region proposal network) and recognition using CNNs. I achieved a mAP (mean average precision) of 0.56.

I developed an inventory optimization platform that used time-series models for long-term forecasting and demand sensing for short-term forecasts. The combined forecast was more accurate and robust for real-world events like weather changes and customer behavior. It supported optimal demand forecasting model selection from many models like LSTM, ARIMA, Holt-Winters, and hybrid models such as AdaBoost-LSTM ensemble.

This helped the client maintain optimal inventory for all products and plan demand fulfillment.

I built a proof-of-concept IoT streaming analytics platform on an 8-node Spark cluster. The stream of real-time sensor data was collected using Kafka, processed using ML models on Spark, and output fed to the live Power BI dashboard.

Project to compare the performance of three different families of AI algorithms for autonomous car parking tasks.

• Reinforcement learning – soft actor-critic (SAC)
• Imitation learning – behavior cloning
• Neuroevolution – genetic algorithm

The above algorithms were used to build three different AI agents that tried to control the steering and acceleration to park the vehicle correctly. The performance of these agents was compared for multiple simulations.

Developed a deep learning model to detect damages in a picture of the car. Implemented Facebook's Detectron model using PyTorch to build an image segmentation model.

Given a pic of the damaged car, find which part is damaged. The parts can be either the rear bumper, front bumper, headlamp, door, or hood.