Pranjal Singh

Designed and developed a framework to simplify the 3-Party Model of Doctor, Patient and Pharmacist in Health domain making their life simpler to deal with day to day health issues.
Real Time Payment using VISA API enabled instantaneous transfer of money between all the parties
Real Time Social Network Feed helped patients and doctors to monitor diseases spread in their respective areas
Technology: Java Spring, Android, SQL, Python, Image Processing, Natural Language… Show more

Designed and developed a framework to simplify the 3-Party Model of Doctor, Patient and Pharmacist in Health domain making their life simpler to deal with day to day health issues.
Real Time Payment using VISA API enabled instantaneous transfer of money between all the parties
Real Time Social Network Feed helped patients and doctors to monitor diseases spread in their respective areas
Technology: Java Spring, Android, SQL, Python, Image Processing, Natural Language Processing
Business Value: Operating time of all parties has a significant impact whereas ease of doing business becomes high Show less

Designed and developed a FP-Tree model to clean raw merchant names coming from transactions in real-time
The model had the ability to identify the junk in the raw name and clean it to give richer information
We achieved a sustainable and intelligent solution that can eliminate the contextual junk from the string using the frequent patterns tree of merchant names
Technology: Hadoop, Spark, Apache Parquet, Hive, Python

Our app focuses on eliminating these challenges via push payments at the predefined toll booth making them self-serviceable resulting in a hassle free automated solution to the toll problem. The authorities get a transparent system to manage fund collections.
We used card to card push payment using Two Way SSL Authentication for a hassle free payment. We used Google Cloud Messaging server to send push messages of successful transaction to toll plaza and upstream message from toll plaza to… Show more

Our app focuses on eliminating these challenges via push payments at the predefined toll booth making them self-serviceable resulting in a hassle free automated solution to the toll problem. The authorities get a transparent system to manage fund collections.
We used card to card push payment using Two Way SSL Authentication for a hassle free payment. We used Google Cloud Messaging server to send push messages of successful transaction to toll plaza and upstream message from toll plaza to send the QR Code to the client. We crawled web using python to collect toll data. Platform is Android
We collected proper and full data from Web using a self-built automated script. We remove the use of cash from the whole ecosystem to make the payment transparent to the financial authorities
Technologies used were Android, Google Cloud Messaging, Python, VISA Push Payment, Zomato API
Business Value: Operating time as well as cost becomes significantly less while there is an extra revenue scope for all parties Show less

Internal productivity tool using LLM with RAG

Designed and developed a predictive model to flag non-compliant transactions which is scalable and high performing
Involved fuzzy string matching to validate string entities and regression to assign feature weights for scoring
Involved Named Entity Recognition using Condition Random Fields for automating validation processes
Technology: Hadoop, Spark, Hive, Java, Spark SQL, Parquet
Business Impact: The framework helped the compliance teams to prevent revenue leakage in order of… Show more

Designed and developed a predictive model to flag non-compliant transactions which is scalable and high performing
Involved fuzzy string matching to validate string entities and regression to assign feature weights for scoring
Involved Named Entity Recognition using Condition Random Fields for automating validation processes
Technology: Hadoop, Spark, Hive, Java, Spark SQL, Parquet
Business Impact: The framework helped the compliance teams to prevent revenue leakage in order of millions due to ambiguities in incoming data
Show less

Analyzed and designed a framework which utilizes k-means clustering to form clusters of redundant records
Hierarchical approach which involved pre-processing followed by unsupervised clustering
Technology: Hadoop, Spark, Hive, Java
Business Impact: The framework reduced the master cache by more than 85% which benefited all downstream systems and improved performance by 35%

Designed and developed a framework to link transactions to an external database using text mining techniques
Adopted to fuzzy string matching techniques for string entities
Technology: Hadoop, Spark, Java, Hive
Business Impact: The model helped in automation of the process saving a big chunk of manual effort as well as improved the quality and quantity of true matches

Designed and developed a real-time merchant stamping framework in Spark for faster downstream consumption
Intelligently reduced lookup size and optimized record matching time with a one-way fingerprint for each transaction
Technology: Hadoop, Spark, Kafka, Java, Python
Business Impact: We achieved a record rate of stamping which had a direct impact on loyalty platforms and helped in easy migration of existing system to LTA

Developed a model and a framework to find merchants based on transaction data (Distorted Merchant Names and other information)
Derived NLP based features and developed Machine Learning based methodology to predict the most matched name from a huge data set of merchant names using intelligent techniques.
Designed a separate module for address matching where we used USPS codes and other location based services to standardize the address and matched based on location based features with a… Show more

Developed a model and a framework to find merchants based on transaction data (Distorted Merchant Names and other information)
Derived NLP based features and developed Machine Learning based methodology to predict the most matched name from a huge data set of merchant names using intelligent techniques.
Designed a separate module for address matching where we used USPS codes and other location based services to standardize the address and matched based on location based features with a weighted model.
Built a Boosting Model of three classifiers that had a significant impact on the accuracy of the model
Accuracy obtained by this hybrid model is ~95.9%
Technology: Hadoop, Spark, Hive, Apache Parquet, Spark SQL, Machine Learning, NLP, Python
Business Impact: This helped us to identify merchants from a multi-million dataset more intelligently and accurately directly impacting systems which throw real-time offers Show less

To identify VISA Merchants' Location Credibility for reducing fraud and risk in transactions
Designed a framework which models merchants and deduce their properties using a scalable and robust Machine Learning model
Used an ensemble of classifiers and DBSCAN clustering to detect anomaly in real-time
Achieved an accuracy of 98% on more than 2 million transactions
Technology: Hadoop, Python, Hive, Spark
Business Impact: Helped in saving millions of transactions from fraud without… Show more

To identify VISA Merchants' Location Credibility for reducing fraud and risk in transactions
Designed a framework which models merchants and deduce their properties using a scalable and robust Machine Learning model
Used an ensemble of classifiers and DBSCAN clustering to detect anomaly in real-time
Achieved an accuracy of 98% on more than 2 million transactions
Technology: Hadoop, Python, Hive, Spark
Business Impact: Helped in saving millions of transactions from fraud without using any personal information Show less

Designed and developed an engine to detect web presence VISA merchants and hence create an offline merchant repository
Location attributes will provide an easy and comprehensive database for merchant address lookups
Technology: Python, Hive
Business Impact: Potential to save millions as it is a self-sufficient merchant finder as well as merchant data linker

We developed a simple and robust technique to generate document vector using decompositional semantics
We proposed a graded weighting schema for building document vectors from word vectors with distributed semantics
We release the largest movie review corpus in Hindi containing 697 movie reviews collected from Dainik Jagran and Navbharat
Achieved state-of-the art results for sentiment analysis on Hindi movie and product reviews, 90.30% & 92.89% accuracies resp.
Achieved state-of-the… Show more

We developed a simple and robust technique to generate document vector using decompositional semantics
We proposed a graded weighting schema for building document vectors from word vectors with distributed semantics
We release the largest movie review corpus in Hindi containing 697 movie reviews collected from Dainik Jagran and Navbharat
Achieved state-of-the art results for sentiment analysis on Hindi movie and product reviews, 90.30% & 92.89% accuracies resp.
Achieved state-of-the art results for sentiment analysis on IMDB movie reviews, 94.19% accuracy, with 1.6% improvement and on Amazon product reviews, 92.91% accuracy, with 7% improvement
Show less

This work presents and discusses several experiments in authorship authentication of short social network postings, an average of 20.6 words, from Facebook. The goal of this research is to determine the degree to which such postings can be authenticated as coming from the purported user and not an intruder. Various sets of stylometry and ad hoc social networking features were developed to categorize short messages from thirty Facebook authors as authentic or non-authentic using Support Vector… Show more

This work presents and discusses several experiments in authorship authentication of short social network postings, an average of 20.6 words, from Facebook. The goal of this research is to determine the degree to which such postings can be authenticated as coming from the purported user and not an intruder. Various sets of stylometry and ad hoc social networking features were developed to categorize short messages from thirty Facebook authors as authentic or non-authentic using Support Vector Machines. The challenges of applying traditional stylometry on short messages were discussed. The test results showed the impact of sample size, features, and user writing style on the effectiveness of authorship authentication, indicating varying degrees of success comparing to previous studies in authorship authentication Show less

Created three dimensional models from a set of images
Used Camera calibration, depth determination, triangulation in the process
Used SIFT descriptors for feature matching
Implementation done in OpenCV-Python

Designed an android app that could control a laptop using Wi-Fi
A laptop/server side java program that manipulates the instructions received over Bluetooth from android device
Server Side script runs a java-robot which could handle mouse movement and keyboard inputs
Android app can handle multiple gestures such as different types of swipe, double tap, multi-touch, etc.
Communication takes over Wi-Fi between laptop and android device

Studied various papers in the area to propose a model that could do unsupervised cross lingual alignment
Proposed a framework that could align cross lingual bitext without supervision
Worked on a manually built corpus in 2 languages (English & Hindi) containing around 50,000 tokens
Used PLSA and ADIOS to provide a base system to the framework
Framework was built by composing work from different papers together
Framework also uses semantic information using WordNet for… Show more

Studied various papers in the area to propose a model that could do unsupervised cross lingual alignment
Proposed a framework that could align cross lingual bitext without supervision
Worked on a manually built corpus in 2 languages (English & Hindi) containing around 50,000 tokens
Used PLSA and ADIOS to provide a base system to the framework
Framework was built by composing work from different papers together
Framework also uses semantic information using WordNet for alignment
Evaluation parameters are precision, recall and F-score Show less

Inthiswork,weexperimentwiththreeapproachestoextractrelationshipbetweencross-lingual clusters and sentences. While other work have mainly focused on few features such as predicates and have not directly looked into the whole document, we consider the whole document and not merely look at few features. Moreover, our work is more on a focused data,i.e. data concentrated on very few topics. We have used unsupervised learning of Natural Languages and concentrated on Hindi and English, though it can… Show more

Inthiswork,weexperimentwiththreeapproachestoextractrelationshipbetweencross-lingual clusters and sentences. While other work have mainly focused on few features such as predicates and have not directly looked into the whole document, we consider the whole document and not merely look at few features. Moreover, our work is more on a focused data,i.e. data concentrated on very few topics. We have used unsupervised learning of Natural Languages and concentrated on Hindi and English, though it can be extended to any language.
We have used unsupervised methods of learning due to their advantage over supervised methods. They do not require an annotated corpus which proves to be quite costly. Also the corpus used has been manually built by us which includes a Hindi corpus and an English corpus containing articles on Coal Scam. Hindi corpus contains around 35000 tokens whereas English corpus consists of over 50000 tokens. Here each line is treated as a document. Show less

In today's world with ever increasing volume of text resources over internet and digital libraries, organizing these documents has become a practical need. Clustering is an important technique which automatically organizes large number of objects into small number of coherent groups. This leads to efficient and effective use of these documents for information retrieval and other such NLP tasks. Clustering algorithms require a metric to quantify how different two given documents are. This… Show more

In today's world with ever increasing volume of text resources over internet and digital libraries, organizing these documents has become a practical need. Clustering is an important technique which automatically organizes large number of objects into small number of coherent groups. This leads to efficient and effective use of these documents for information retrieval and other such NLP tasks. Clustering algorithms require a metric to quantify how different two given documents are. This difference is often measured by some distance measure such as Euclidean distance, Cosine similarity to name a few. In this work, we experiment with five well known distance measures and compare their performance on seven datasets using k-means clustering algorithm. The work shows that some distance measures clearly outperform others. Also, which measure to use in a specific scenario depends on what type of clusters does one want for the task at hand.
Show less

Recommendation System can provide suggestions about movies, videos, newspaper articles to a user. For instance, they may predict whether a user would like a movie or not. These systems collects information from a large number of users and mainly use nearest neighbor techniques to provide recommendations. In this project, we have built a movie-recommendation system using MovieLens Database. We have tried building and using several systems which mainly use a Machine Learning approach called… Show more

Recommendation System can provide suggestions about movies, videos, newspaper articles to a user. For instance, they may predict whether a user would like a movie or not. These systems collects information from a large number of users and mainly use nearest neighbor techniques to provide recommendations. In this project, we have built a movie-recommendation system using MovieLens Database. We have tried building and using several systems which mainly use a Machine Learning approach called Collaborative filtering. We have tried both the approaches, Item Based and User Based Collaborative filtering and compared the results with the already developed systems.
Show less

One of the biggest trends in wireless communications has been the explosion in demand for data services driven by the introduction of smartphones. Meeting this demand requires large investments in wireless capacity. However, for various socioeconomic reasons, the price of basic service that a service provider can charge end users is fixed, often at a level that cannot generate enough revenue to pay for the cost of upgrades. Therefore, to stay on a sustainable path, the service provider needs to… Show more

One of the biggest trends in wireless communications has been the explosion in demand for data services driven by the introduction of smartphones. Meeting this demand requires large investments in wireless capacity. However, for various socioeconomic reasons, the price of basic service that a service provider can charge end users is fixed, often at a level that cannot generate enough revenue to pay for the cost of upgrades. Therefore, to stay on a sustainable path, the service provider needs to explore income from other sources, preferably in a mutually beneficial manner. The motivation is that content provider wants maximum number of hits, user wants to access maximum data at same price and operator wants to earn more revenue without disturbing the other two. We consider a solution that allows the content provider to “sponsor” its content so that it does not get charged to the end users’ monthly quotas. The arrangement removes end users’ concern about paying an uncertain amount of bandwidth cost for carrying advertisements. As a consequence, more content will be accessed, not only because some of it is free but also because users are effectively given more quota. The content provider’s profit increases as long as the cost of sponsoring stays below the new advertising revenue from increased viewing of its content. Content sponsoring also benefits the service provider by giving it the opportunity to charge content providers who, have a greater willingness to pay than end users. Income from this new source enables the service provider to recover the value of some of the mobile services that it is enabling and use that revenue to finance capacity expansion. We mimic the system by setting a VPN server and a nodeJS server-script which records user activity and serve free contents to user by manipulating incoming data. Our Android App shows free stuff specific to each user and also shows them their usage details.
Technology: MySQL, VPN Server, NodeJS, Android Show less

Implemented a working compiler for programming language ADA(ver2). Applied various aspects of a compiler learned in the course.
Implemented Lexer and Parser in C using LEX and YACC along with Semantic Rules for Symbol Table and used 3-Address Code as IR which is then translated into Assembly Language for MIPS

Built a web application to make the process of Hostel Management electronic
Used Entity-Relationship modeling, normalization to Boyce Codd Normal Form (BCNF), and integrity constraints to design the database
Implemented the portal having facilities for creating students profile, staff profile, complaint addressing etc.
Interface programming was done in HTML and JavaScript, database built on MySQL and the interaction between them was done by PHP hosted on Apache server
Built a web… Show more

Built a web application to make the process of Hostel Management electronic
Used Entity-Relationship modeling, normalization to Boyce Codd Normal Form (BCNF), and integrity constraints to design the database
Implemented the portal having facilities for creating students profile, staff profile, complaint addressing etc.
Interface programming was done in HTML and JavaScript, database built on MySQL and the interaction between them was done by PHP hosted on Apache server
Built a web application with Apache/PHP as the frontend and MySQL as the backend Show less

Implemented packet sniffer and tested it on hostel LAN network
Implementation was done in Python and used PCAP Library using sockets
Captured almost all the packets in the local network by acting as a gateway
Can log all TCP, HTTP and UDP data over the network, show values of various fields in the packet and monitor network usage

Implementation of System Calls through various Posix functions.
Implementation of Posix message queues and threads, Virtual Memory, File System and User Processes

Implemented an interpreter to take an Abstract Syntax Tree (AST) of Oz and output the execution steps
Included implementation of Single Assignment Store, semantic stack and closure
In the Mozart OZ environment, developed an interpreter for a declarative sequential language
Implemented a Semantic Stack and the Single Assignment store to output the sequence of execution states given a source code as an input
Implemented procedure definitions and calls, Scoping and… Show more

Implemented an interpreter to take an Abstract Syntax Tree (AST) of Oz and output the execution steps
Included implementation of Single Assignment Store, semantic stack and closure
In the Mozart OZ environment, developed an interpreter for a declarative sequential language
Implemented a Semantic Stack and the Single Assignment store to output the sequence of execution states given a source code as an input
Implemented procedure definitions and calls, Scoping and conditions
Developed an interpreter for a declarative sequential language in Mozart Oz environment
Implemented procedure definitions and calls, scoping and conditions
Show less

In the Multiway Cut problem, we are given an edge-weighted graph, a subset of the vertices called terminals and a cost k, and asked for a k - weight set of edges that separates each terminal from all the others. When the number n of terminals is two, this is simply the min-cut, max-flow problem, and can be solved in polynomial time. The problem becomes NP-hard as soon as n = 3, but can be solved in polynomial time for planar graphs for any fixed n. The planar problem is NP-hard, however, if n is… Show more

In the Multiway Cut problem, we are given an edge-weighted graph, a subset of the vertices called terminals and a cost k, and asked for a k - weight set of edges that separates each terminal from all the others. When the number n of terminals is two, this is simply the min-cut, max-flow problem, and can be solved in polynomial time. The problem becomes NP-hard as soon as n = 3, but can be solved in polynomial time for planar graphs for any fixed n. The planar problem is NP-hard, however, if n is not fixed. Show less

Emotions have been widely studied in psychology and behavior sciences, as they are an impor- tant element of human nature. In this project we aim to identify certain basic emotions present in a text. We have used ISEAR corpus which contains sentences related to certain emotions, a PCFG parser and determind the weight of a sentiment and then tried to give weight to context with the help of PLSA algorithm which is basically a probabilistic method to determine the meaning behind the words. The… Show more

Emotions have been widely studied in psychology and behavior sciences, as they are an impor- tant element of human nature. In this project we aim to identify certain basic emotions present in a text. We have used ISEAR corpus which contains sentences related to certain emotions, a PCFG parser and determind the weight of a sentiment and then tried to give weight to context with the help of PLSA algorithm which is basically a probabilistic method to determine the meaning behind the words. The results of the project are quite promising.
Show less

--Used Supervised and Unsupervised training.
--PCFG parser was used for parsing the sentences, and the tokens were searched for in the emotion tagged dictionary.
--In the Unsupervised way, PLSA (Probabilistic Latent Semantic Analysis) algorithm was used to get the probabilities for the training data. The sentence was then folded to get the probability of the required emotion category.
--Implemented in Python and C

Programmed a 32-bit ALU on a XILINX Spartan3 FPGA Board, using BSV as hardware programming language
Implemented and designed a simple processor on a FPGA. In particular implemented an ALU, and Register File on a XILINX SPARTAN 3 FPGA board
Implemented CPU and CLOCK on Field Programmable Gate Arrays FPGA, using Verilog as HDL
The processor was capable of taking 2 4-bit inputs and performing mathematical operations on them
Input was fed through switches and output shown as 7-Segment… Show more

Programmed a 32-bit ALU on a XILINX Spartan3 FPGA Board, using BSV as hardware programming language
Implemented and designed a simple processor on a FPGA. In particular implemented an ALU, and Register File on a XILINX SPARTAN 3 FPGA board
Implemented CPU and CLOCK on Field Programmable Gate Arrays FPGA, using Verilog as HDL
The processor was capable of taking 2 4-bit inputs and performing mathematical operations on them
Input was fed through switches and output shown as 7-Segment LED Display on the FPGA.
Implemented a digital clock. The clock had 4 modes which were: HH:MM mode, MM:SS mode, Alarm clock mode i.e. LED glowed when the time reaches the alarm time for a min, and Stopwatch mode having options to start, reset and stop
Implemented ALU, Processor and Register File. 4-bit Input fed through switches and output shown as 7-Segment LED Display on the FPGA Show less