Sayan Mukhopadhyay

Constructing a framework to validate machine learning models and data

• ReverseAd Keyword Algorithm System
• Whizhack Anomaly Detection System - developed the product from scratch, including the REST API and ML model
• Sulvo Ad Price Predictor System – Sulvo

The system is comprised of three main components: a collector, a real-time trainer, and a predictor. The collector listens to programmatic ad selling information for each impression and sends the predicted floor price for each impression to the ad server through the predictor. The real-time trainer builds a model using recently collected data and saves it, while the predictor uses the model for predictions. The Sulvo ad server is hosted on AWS cloud, while the real-time trainer and predictor are hosted on Google cloud platform. The collector, which is also hosted on AWS and exposes a REST API to collect data, receives information from the ad server and pushes it to a Redis message queue. When the message is dequeued from the Redis server, it is pushed to a Google Big Query instance. The real-time trainer, hosted on a Google Compute Engine, fetches the latest information from Big Query and builds a multi-layer CNN model using TensorFlow, saving the model in Google Data Store. The predictor server, a Falcon-based REST API hosted on Google App Engine, receives prediction requests from the Sulvo ad server and returns a floor value calculated based on data stored in the Data Store. This system is deep learning-based and is able to train the model with real-time data, providing high performance with a response time of under 300ms to the ad server and the ability to handle up to 22 million predictions daily. My responsibilities included building the product from scratch, from the data collector to the ML model.

Future Today Recommender is designed to update the stored model in each training iteration rather than replacing it. This allows for more accurate recommendations over time. The model is made accessible through a Rest API, which allows it to be integrated with other systems. To ensure fast response times, the API stores all necessary information for recommendations in shared data structures in RAM. The training process is scalable and can handle large input data sizes, even those larger than the size of the system's RAM. It is implemented using Spark and is able to handle a high volume of user views, up to 40 million per month, with a response time of under 150ms for prediction requests. I was responsible for building products from scratch, which included developing both the REST API and the ML model.

I worked on a project for a real estate client to automate manual reporting processes detailing data for High-Net-Worth clients. I developed a two-step approach that involved gathering relevant URLs through a Tropical Crawling approach and then extracting names, organizations, and locations from the output URLs using a name entity recognition algorithm. The Tropical Crawling approach utilized two sets of URLs as input: seed URLs where the crawling would begin, and target URLs as reference URLs and a known subset of relevant URLs. I used Jaccard Distance as a similarity measure to compare the content of new URLs with the targets and select the most similar ones. This process was repeated, using the selected URLs as seed URLs and considering all output URLs from previous steps while excluding those that had already been selected. To reduce the search space, I established a threshold in the similarity measure for URLs, using the average similarity among the target URLs. I also created a set of keywords, such as "profile" and "biography", for relevant URL strings. The Name Entity Recognition algorithm enabled the extraction of names, organizations, and locations from the selected output URL contents using the Beautifulsoup library to parse HTML pages and the Stanford NER NLTK library. My responsibilities included formulating the problem, designing the solution architecture, mentoring implementation, and serving as a client liaison.

I worked on a project to recommend the best vendors for a particular proa duct ID using purchase order and invoice information. To solve this problem, I divided it into two sub-areas. The first involved predicting the probability of successful purchases. I used three approaches for this calculation: a regression-based approach using Logistic Regression, a rule-based approach using the Random Forest Algorithm, and a Bayes-theory approach using a Naive-based classifier. In the second phase, I grouped the data by vendor ID (VEN_ID) and calculated each vendor's average probability of a successful purchase. I then sorted the vendors according to this probability and applied rule-based engines, such as selecting the vendor with the minimum delivery time or the maximum number of deliveries, to recommend the most suitable vendor. I prioritized the rules by predicting user query aspects using classification. My responsibilities included formulating the problem, designing the solution architecture, mentoring implementation, and serving as a client liaison.

I applied a text classification algorithm in brand reputation and promotion management, using various textual data from the web, CRM, and mail servers. This data was collected through a crawler script or third-party API and stored in HDFS. I classified sentences according to the degree of sentiment (positive, negative, neutral) and by brand dimension (food, staff, entertainment), providing trend summaries and KPI information to data brand owners and market analysts. I selected the naive Bayes algorithm as the text classifier following a literature survey, implemented it using Mahout, and visualized the big data using Kibana, keeping the data in elastic search. My responsibilities included requirement analysis, application architecture design, database schema design, unit testing, integrated testing, planning for the release strategy, and creating a proof-of-concept.

I developed a generic framework for estimating and forecasting quantitative features from a large historical data set (100 million orders per day, supporting 3 months of data). My proprietary implementation of machine learning methods allowed me to realize quantitative target features (number of impressions, ECPM) as a linear regression of a set of base features (geo, site, ad size, frequency). I applied linear regression to implement the least squares method. The overall average accuracy improved by 60%+ and selected premium sites in smaller data sizes had an accuracy of 80%. My responsibilities included requirement analysis, application architecture design, database schema design, unit testing, integrated testing, planning for the release strategy, and creating a proof-of-concept.

Developed a virtual order management system and simulated it in a high transaction environment. The system was instrumented using ARM, MY-ARM, and Kronos API, and the performance characteristics, such as accuracy, instrumentation overload in execution time, memory, and processor usage were compared. The log file was parsed in real-time and a distributed, divide-and-conquer architecture was implemented for the Real-time Log-file Parsing Framework. Data of potential interest was transmitted to a central server through a robust and highly tuneable transport layer, and heavy-weight parsing was completed in the central server separate from the production host. Order latency was divided into two phases (inside and outside the box) to enable further analysis. Inside-the-box latency was measured using a Kernel program and outside-the-box latency was collected using the Corvil API and WSDL.

My responsibilities included design, development, performance testing, and production release of a Socat-based shell script and a C++-based log development.

I worked on an application support project to modify and enhance an in-house position management, portfolio valuation, and risk summary product. The application pulled market data through Reuters API and trade information from Fidesa, calculated various portfolio valuations and risk analysis parameters, and visualized the information for traders. My responsibilities included front/back-end design, development, unit testing and profiling, beta calculation, average volume calculation, industry-wise index decomposition, and exception management. I was also part of a risk analysis group member (Imagine) of CSFB Prop-IT.

• Developed a tool to monitor network, server, and application alerts for all devices used by prop traders
• Front end of the tool is a grid view made with WPF application
• Back ends are different adapters that run in parallel threads and retrieve data from various sources using WSDL and telnet clients
• Retrieves data on application alerts from a web server through log-file parsing
• Device names and connected devices are retrieved by querying a configuration database
• Project was for an in-house enterprise product of Credit-Suisse
• Responsibilities included designing and developing the tool, unit testing and profiling data adapter parts, and creating a front end explorer view connected to a database through Xml Document.