Deekshith B

Deep Learning for forecasting

• Lead evaluation of deep learning models for E-commerce product demand forecasting which replaced traditional models being used and became overall forecasting approach for Flipkart demand planning team
• Built Deep learning experimental framework and implemented current SOTA deep learning techniques including Seq2Seq with attention variants, Temporal Fusion Transformer, MQRNN, TSMixer including making architectural changes resulting in overall 5% wMape improvement across business units

Private Label Forecasting

• Built and productionised ensemble of Boosting, ARIMA, ETS models to predict long lead time(>90 days) demand forecast for Flipkart's privately owned brands with improvements of over 25% wMape compared to baseline.

RCPV Mishipment returns fraud detection

• Worked as a part of Flipkart's Trust and safety team on Fraud detection as part of which work included Brainstorming potential features, Performing EDA, feature engineering, and did modeling using gradient boosting machines and was able to increase AUC by 2.3% over the baseline model.

Supply planning network optimisation

• Built and productionised an optimization model using cvxopt that creates a supply network of how shipments will flow from source(warehouse) to destination based on business inputs which is then used by supply planning for better Manpower planning, line haul planning resulting in SLA savings.
• Came up with the idea and optimized the existing model and reduced runtime by 8 hours by using parallelization and Divide & conquer techniques.

Supply planning vendor allocation

• Brainstormed requirements with business users and has designed and formulated an optimisation problem of vendor selection and quantity allocation with cost minimisation as objective. Mentored a Data scientist to implement it, which resulted in 5 million USD/year in savings.

Lip Reading using Deep Learning

• Implemented audio-visual, encoder-decoder deep learning model that predicts words in a sentence from a video clip, the architecture is based on the paper “Lip reading sentences in the wild” by Andrew Zisserman and others, the model was implemented in Keras and trained on the GRID-corpus dataset

Face Recognition for Surveillance

• Built a Face Recognition program targeted at surveillance products at Collins Aerospace based on the paper “FaceNet: A Unified Embedding for Face Recognition and Clustering”, we used online reference material to fit the requirements of the problem we were tackling, such as real-time training, low false-positive identification, etc