Trained a random forest regressor to predict potential holdings for a fund and achieved 20% better returns vs historical performance.

Trained a CNN to produce simultaneous localization and mapping using sensor fusion technology 83% accuracy.

Automated leasing communication between prospect and clients of a real estate company through MLaaS model, which resulted in an increase of 3.5% occupancy.

Developed MLaaS model to automate inter-company processes of a supply chain that reduced delivery delays by 13%.

Created an Automated Number Plate Recognition system using CNNs, able to detect number plates with 95% accuracy.

Task: designing a machine learning model to automate leasing communication between clients and prospects of a real estate company

The challenge was increasing the (P/V)profit to volume ratio of the company while being compliant with the existing legal housing and leasing norms.This required extensive hypothesis gathering from company's finance department and legal department,the new gathered data was then merged existing data from client and prospect communication

In order to resolve such hetrogenous data,we developed a deep neural network, 2 different LGBM models and a NLP model were built to calculate credit score,sentiment etc and ensembled to give a custom prospect score.

The prospects were prioritized as per the score and followed up accordingly.The model was able to improve the occupancy by 3.5%.

Challenge: creating a drone which could simultaneously map it's environment(terrain objects etc) and localize each different objects in its environment with precision.The purpose of this drone is providing assistance in mining industry,construction industry(for instance: exploring difficult terrains,ensuring work safety etc.) .

The drone is equipped with a camera and some other GIS sensors(gyroscope,lidar,etc) which gathers the data for being analyzed .

The Initial machine learning model developed for this task consisted of a CNN bottle neck architecture + RNN but this had some issues- dead neurons at the bottleneck due to uneven light,high temperature environment or inclement weather conditions(ie- not able to properly localize objects in the environment),high memory access cost(as this CNN architecture had lot of computation,it tended to be slow).

To address this problem a we used simpler CNN model (a variant Mobilenet V2 model with seq-squeeze blocks) which was faster and lighter than previous model + a Variational Autoencoder to process latent variables(temprature,weather etc)+RNN to deal with time variant factors