Product @ Mastercard | Harvard PAIR | IIM C-IIT Kgp-ISI K | ISB YLP | BITS Pilani

Aditya Gadepalli (AG)  • Peerlist

Studied the application of Phase Change Materials in Building materials to reduce cooling load & primary energy usage for various geographies. 

Advanced Materials for Energy Efficient Buildings

Developed an ensemble of machine learning classifiers to predict the Click Through Rate (CTR) based upon multiple parameters obtained from a click stream history. 

We used Tensorflow 2.0 to build a Deep Neural Net based Linear Combined classifier to achieve "Memorization" and "Generalization" simultaneously. The results were bench marked against the standard Machine Learning models like logistic regression, Gaussian Naive Bayes, Random forest, KNN, XGBoost and AdaBoost based classifiers.

Click Through Rate Prediction using Wide and Deep Networks

We forecast the future trajectories associated with COVID-19 for the coming days in India using Auto Regressive Integrated Moving Average model (ARIMA), Curve (Spline) fitting, Logit Model and modified SIR (Susceptible, Infected and Removed) based pandemic model. Also, we compare these models and explain the necessary conditions that need to be satisfied (via normalization techniques) to minimize modelling violations. The short term forecasts obtained so far have been highly accurate.


Comprehending COVID-19 : A Time Series and Pandemic Modellin

Led a team of 8 students to survey & review the appraisal systems of NTPC and IBM to measure its influence on employee’s work outputs.

Critical analysis of Performance appraisal systems

Identified and optimized the control parameters to achieve an ISO 5 level cleanliness environment in the cleanroom on campus. Performed CFD simulations on COMSOL 5.1 for particle flow analysis and used Design of Experiments(DOE) methods to optimize test cases. Finally, we validated the obtained results along with Hemair Systems India Ltd.

Design of Cleanroom for Micromachine(MEMS) Fabrication

We developed a computationally lighter segmentation architecture based on U-Net which utilizes Simple Canny Edge Detection to enhance edge detection to preserve geometric snapping of the predicted segmentation masks. Yet to validate our results on all standard datasets against the some state-of-the-art models.

Edge Enhanced U-Net for Aerial Image Segmentation of Buildin

Led a team of 8 students to design and build the product from scratch and achieved 20 micron erosion on AISI 1020 steel using brass electrode.

Fabrication of portable Electrodischarge Machining Machine

In an attempt to delve into the field on Ethical AI, this work is an inspiration from the July 2019 paper published by LinkedIn USA. We work on implementing the four bias mitigation algorithms proposed by the authors using synthetically generated data.

Fairness-Aware Ranking in Search & Recommendation Systems

We propose a “Hybrid Hyperlocal” model which combines the strength of both the traditional warehouse-inventory model and the much recent Hyperlocal model to achieve a category specific distribution mechanism for the e-commerce giants.  Using the principles of Game Theory we also propose simultaneous growth models for all member of the ecosystem whilst also leveraging the new Hybrid mechanism to reduce "Stress" over conventional supply chains to achieve better profitability.

Hybrid Hyperlocal Delivery Systems

Formulated algorithms and built a GUI using Matlab to obtain splines and their respective blending functions for any given dynamic input.

Interactive Creation of Splines

In the age of information, there is a growing need to process data, both structured and unstructured, both with ease and accuracy. Such an intensive task might be achieved through machine learning based applications but not always. Yet they can be made possible through collaborative yet statistically designed efforts like “crowdsourcing” wherein the task at hand is broken into smaller “micro-tasks” usually needing binary decisions to be exercised by the amateur human participants working upon them. The use of such collective intelligence can simplify few of the difficult database management tasks. One such utility is found with respect to identifying the functional dependencies within (large) datasets whilst maintaining data privacy all throughout. Moreover, when such activities can be further designed into game-like interfaces, there is an increased pace and motivation within the crowd (workers) executing the task(s).

Scalable Database Normalization powered by the Crowd

Consumer facing organizations need to constantly keep track of feedback/grievances from their stakeholders. Relevant feedback must be classified and incumbent issues need to be clustered into appropriate categories. This clustering needs to be assessed over several indicators and benchmarks that convey effect over improvements in functional and operational utility. We propose a semi-supervised approach for performing relevant news article selection using extensive text cleaning, pre-trained RoBERTa embedding and Latent Semantic Analysis, followed by extensive exploratory data analysis, complex network analysis and sentiment extraction on the developed corpus. The processed articles are subjected to Latent Dirichlet Allocation, Hierarchical Dirichlet Process and Clustering to discover potential topics in an unsupervised manner. We then label the articles accordingly to devise a training set for building a multi-class classifier for topic categorization and assignment. The Top-3 accuracy metric is chosen along with several other indicators to understand the performance of the classifier and also ascertain the severity and other aspects of the incoming news articles. New topic discovery is also performed to improve the classifier classes and the model is re-calibrated at appropriate intervals to stay above human benchmark in scale and accuracy.