Builder | AI | ML

Chirag Tailor  • Peerlist

- Implementation of a Random Forest classifier with Decision Trees for the UCI wine quality dataset.
- Implementation of a Naive Bayes Classifier in Python for UCI mushroom dataset.
- Automated stock trader using Decision Trees with Boosting and Q-Learning agents.
- Developed an Isolation Game playing AI using iterative deepening and Alpha-Beta pruning.
- Formulated the N-Queens problem into an optimization problem and implemented a Genetic Algorithm solution.


Machine Learning Projects

- Developed a market simulator to execute artificial option trades and analyze portfolio performance
using metrics such as Sharpe Ratio, Sortino Ratio, and WIn/Loss Rate.
- Analyzed the performance of common option trading strategies such as Credit/Debit Spreads, Iron Condor, and Long/Short Straddles.
- Developing a machine learning algorithm to optimally choose Credit Spread width, time horizon, and exit strategy.

Option Strategy Backtester

-	PARQR is an ongoing Master’s research project in collaboration with Dr. Thad Starner and funded by Zvi Galil, the Dean of Georgia Tech College of Computing.
-	Leading a team of graduate researchers and software engineers with the objective of using AI and Natural Language Processing to help students find answers to their questions on Piazza faster and leverage student usage data to provide actionable insight to MOOC Professors.
-	Designed a dockerized, horizontally-scalable Redis queue and worker architecture to parse latest post information and train new recommendation models. Recommendations are served from a RESTful Flask API built on a MongoDB store. Working on developing an A/B testing framework.
-	Experimenting with improvements to traditional TF-IDF document similarity with incorporation of PageRank or replacing BOW vectors with distributed BOW vectors (dov2vec). 
-	Available for download on Chrome Web Store.



Piazza Automatic Related Question Recommender (PARQR)

- Master’s Project in collaboration with Dr. Polo Chau and a team of graduate students.
- Designed a cluster of over 30 machines to reliably detect anomalies in power sensor data at scale.
- Built a data ingestion pipeline using a Java multithreaded reverse proxy to write data to HBase.
- Created a Spark Streaming analysis pipeline to perform anomaly detection at scale.
- Results: Ingestion speeds of 400k data points per second on a 30 node cluster with linear scalability.
- Published in International Parallel & Distributed Processing Symposium Workshop.


Scalable Architecture and Visualization for Anomaly Detectio

- Prototyping an ASR pipeline with Librosa for speech processing and Keras + Tensorflow for building CNN architecture.
- Experimenting with various speech-to-image feature extraction methods including log-melspectrogram + ∆'s + ∆∆'s, posterior confidence calculation methods, and CNN architectures with low false-accept and false-reject rates but also minimal parameters and multiplies.
- Challenges include scarcity of labelled training data, building a network robust to noise and reverberation, and producing fast predictions with low latency. 


Streaming Spoken Keyword Spotting using Convolutional Neural

- A user-to-user collaborative filtering system augmented with information of user’s friend lists from social media platforms. The system also discovered and provided preferential treatment for “foodies”.
- Implemented on AWS EMR with PySpark for distributed analytics and Apache Hive for data storage.
