I'm Naveen Chakravarthy Balasubramanian.

Forecasting Cost of Sales

  Forecasting the cost of sales incurred in the sale of a product, in terms of enlistment of services or procurement of equipment. Built the process from the ground up and continuously improved it, till it became the gold standard in the segment. The forecasting is done at two different granularities, to create value to both the individual segments and higher management - detailed level and a bird's eye view. This reduced the lead time from 2 days to 1 hour and reduced the annual effort by 84 FTEs across the world. The MAPE is consistently less than 7%. End to end orchestration using pipelines enabled frequent planning.

Free Cash Flow forecasting

  Forecasting the free cash flow available as an algebraic sum of inflow and outflow components. The inflow AR and outflow AP were forecasted separately using ARIMA and Exponential Smoothing techniques after validating for stationarity. This helped reduce quite a lot of hassle in the traditional method while enabling a seamless end to end workflow. End to end orchestration using pipelines enabled frequent planning.

Best Channel To Contact

  Operational costs of contacting a customer average around 2000 USD per state per year with the default option being a call. Since the customer consent and activity data was readily available, it was leveraged to find out the best channel to contact a customer, while taking into account the cost of the operation. Whenever there is an almost equal chances of being reached over an email and over a call, it is preferable to contact over an email, since it is way cheaper. Implementation of this optimization involved building a neural network as a multiclass classification model, to predict the best channel. This brought the operational costs down to 300 USD per state per year - a 76% savings.

Lean Prediction of Shear Strength

  As the conventional method of design of a beam is theoretical and arbitrary, a new approach to use Machine Learning to closely predict the Shear Strength of beams under field conditions is being proposed in collaboration with Er. Suganthi Chandrasekaran, Assistant Engineer, Tamilnadu Highways Department. The current approach, though safe, does not give out economic results. It results in an oversafe and costly design. This proposed method, while ensuring the serviceability and shear capacity of the beam being in the safe zones, gives out an economic design of the beam.

Jarvis

  A question answering system that answers questions about Marvel Cinematic Universe. The model is a transformer-based neural network model that gets trained by taking in question answer pairs. Transformers are state-of-the-art models that leverage the Multi-Head Attention mechanism that puts them above RNNs, LSTMs and GRUs in the respect of sequence processing. This was done as a fun side project with custom made dataset for every movie.

Sudoku Solver

  This is a fun application that looks at an image of a 9x9 Sudoku puzzle and attempts to solve it instantly. This leverages computer vision to extract the puzzle and uses a strategic algorithm to solve the puzzle, rather than the most commonly used Machine Learning Approach. The algorithm currently has an accuracy of 97.5% on the very popular Kaggle Dataset of 1 Million Sudoku Puzzles.