Pages on Ujjwal Chowdhury

Certificates

Mon, 01 Jan 0001 00:00:00 +0000

Microsoft Certified: Azure Data Scientist Associate

In this course, I gained subject matter expertise in applying data science and machine learning to implement and manage machine learning workloads on Azure. I learned how to design and create suitable environments for data science projects, explore and analyze data, and train machine learning models. The course covered the implementation of pipelines and running jobs to prepare models for production. I also acquired skills in managing, deploying, and monitoring scalable machine learning solutions. Throughout the course, I developed a solid understanding and practical experience with Azure Machine Learning and MLflow.

link

Skills Learned:

MLFlow, Docker, Kubernetes, Cloud Computing, AutoML, CI/CD Pipeline, Data Bricks, PySpark, CosmosDB, Blob Storage, Data Visualization.

NISM-Series-XV: Research Analyst

Completing the NISM Series XV Research Analyst examination equipped me with foundational knowledge of the Indian Securities Markets and essential terminologies used in Equity and Debt Markets. I learned diverse Research Methodologies, including top-down and bottom-up approaches, facilitating comprehensive analysis. Additionally, I honed skills in Economic Analysis, Industry and Company Evaluation, Risk Management, Valuation principles, and effective communication through research Report Writing.

link

Skills Learned:

Micro and Macro-economic Analysis, Fundamental Analysis, Industry Analysis, Company Analysis, KPI Analysis.

Artificial Intelligence (AI) for Investments

This program aims to prepare professionals in the finance industry and related fields for the transformative impact of machine learning (ML) and AI. It covers a comprehensive curriculum designed to equip participants with the knowledge and skills necessary to leverage AI and ML techniques in financial markets, trading, and asset management. From understanding financial markets and risk-return analysis to advanced time-series algorithms for risk management, the course provides practical insights and tools for improving investment decisions using AI-based models.

link

Skills Learned:

Portfolio Optimization, CAPM, Algorithmic Trading, ARMA, ARIMA, ARCH, GARCH, 3-Factor Fama-French Model.

Cloud Computing and Distributed Systems

In the course, I learned about Cloud Computing fundamentals, including virtualization techniques such as bare metal and hosted hypervisor models, and the benefits and types of virtualization. Additionally, I gained insights into network virtualization methods, software-defined networking, and inter-data center networking in geo-distributed cloud environments. Furthermore, the course covered topics like leader election algorithms, distributed mutual exclusion, consensus mechanisms like Paxos, recovery approaches in distributed systems, and the design and operation of key-value stores, along with discussions on P2P systems and cloud applications like MapReduce, Spark, and Apache Kafka.

link

Skills Learned:

Virtual Machines, Distributed Systems, NoSQL, MapReduce, PySpark, Apache Kafka.

Data Base Management System

The course on Databases provides a comprehensive understanding of Structured Database Management Systems (DBMS) and their significance in various applications, covering both commercial systems like Oracle and open-source solutions such as MySQL. It delves into data structures, file organizations, SQL, data analysis, database design, and optimization, facilitating proficiency in relational data models, entity-relationship modeling, and query coding practices. Additionally, the course explores multi-tier client/server architectures, web-based database applications, and advanced topics including indexing, hashing, concurrency control, and recovery systems across its eight-week layout.

link

Skills Learned:

RDBMS, SQL, Indexing, Hashing, Query Optimization, window function.

Deep Learning for Computer Vision

This course on Computer Vision begins with an overview of image and video analysis, emphasizing its applications in security, healthcare, entertainment, and mobility. It covers traditional computer vision topics and progresses to deep learning methods, assuming prior knowledge of machine learning. The curriculum includes a comprehensive exploration of visual features, matching techniques, deep learning fundamentals, CNN architectures, RNNs, attention models, generative models, and their applications in real-world scenarios over twelve weeks.

link

Skills Learned:

Classical CV, Multi-layer Perceptrons, Backpropagation, CNN, RNN, AlexNet, VGG, ResNets, U-Net, VAE, GAN, Zero-shot, One-shot, Few-shot Learning.

Data Science Math Skills

This data science course focuses on essential math skills crucial for success in data science, catering to learners with basic math proficiency who may not have studied algebra or pre-calculus. It systematically introduces core mathematical concepts without overwhelming complexity, ensuring learners grasp each idea and math symbol gradually. Upon completion, students will have mastered fundamental vocabulary, notation, concepts, and algebraic rules necessary for advanced data science studies. Topics covered include set theory, properties of real numbers, interval notation, summation, Cartesian plane math, functions and inverses, instantaneous rate of change, exponents, logarithms, and probability theory, including Bayes’ theorem.

link

Skills Learned:

Probability, Bayes Theory, Set Theory, Linear Algebra, Calculus.

Reinforcement Learning

This Reinforcement Learning course offers a comprehensive journey through foundational and advanced RL concepts. It begins with bandit algorithms for single-step decision-making and transitions into full RL with Markov Decision Processes (MDPs) and Bellman Optimality principles for sequential decision problems. Key techniques like Temporal Difference (TD) methods, eligibility traces, and function approximation are explored for value-based learning. Advanced methods, including Deep Q-Networks (DQN), policy gradients, and hierarchical RL, are covered to handle complex, real-world scenarios. The course concludes with Partially Observable MDPs (POMDPs), addressing decision-making in uncertain and incomplete environments.

link

Skills Learned:

Bandit algorithms, MDPs, Bellman Optimality, Eligibility Traces, Hierarchical RL, Policy Gradient for Full RL, POMDPs.

Publications

Mon, 01 Jan 0001 00:00:00 +0000

Investigate How Market Behaves: Toward an Explanatory Multitasking Based Analytical Model for Financial Investments

Sarmistha Das, Ujjwal Chowdhury, NS Lijin, Atulya Deep, Sriparna Saha, Alka Maurya

IEEE Access, vol. 12, pp. 30928–30940, February 2024 · DOI: 10.1109/ACCESS.2024.3369033

Abstract. In recent years, the intersection between financial market behavior and social media has emerged as a sought-after source of information, meeting the requirements of investors, institutions, regulators, researchers, and policymakers. Assessing sentiment and emotions aids in evaluating public psychology on particular stocks, assets, or the overall market, with shifts often aligning with market movements. Previously, machine learning, both traditional and deep learning methods, targeted discerning stock market sentiment and emotion without conducting studies to offer comprehensive explanations for these behavioral factors. In this study, we introduce a multitasking sequence-to-sequence model that integrates financial investment analysis with sentiment and emotion analysis from tweets upheld by an explanation mechanism. We also present the FinEMA dataset, featuring sentiment, emotion, and cause labels on financial stock market changes. Our study highlights how joint learning improves performance in discerning sentiment and emotion by utilizing interrelated features, enhancing task effectiveness. Our proposed model, the Emotion-Sentiment Attention Network (ESAN), achieved 89% accuracy in sentiment identification and 79% accuracy in emotion recognition, outperforming conventional machine learning methods. Furthermore, our findings indicate a positive outlook for the stock market in the latter half of 2023, which has intensified investor optimism, though some individuals still harbor uncertainties. Conclusively, our results suggest that regenerating existing computational tools can open up new research opportunities to address relevant novel tasks. The primary aim of this study is to elucidate the diverse dimensions of financial market behaviour and offer explanatory insights for the research community. The authors maintain impartiality towards specific stocks. It’s essential to note that stock market investments inherently carry market risks and potential losses.

Resume

Mon, 01 Jan 0001 00:00:00 +0000

Download resume (PDF)

Profile

Data Scientist with 3+ years of post-qualification experience developing and implementing ML, Gen AI, NLP and content analysis systems across Banking, Fraud/AML and FinTech. Expertise in Python, R, PySpark, NLP and Gen AI, backed by direct hands-on collaboration with 30+ Indian public and private sector banks to develop and deploy production-grade AI solutions. Proven in AI Governance, model monitoring and end-to-end AI lifecycle management, from AI strategy and technical architecture through validation, deployment and scaling.

Experience

Data Scientist (Fraud & Risk Analytics)

Reserve Bank Innovation Hub (RBIH) · Bengaluru, India · Feb 2025 – Present

Enterprise AI Deployment (MuleHunter.AI): Spearheaded the AI architecture and nationwide deployment of MuleHunter.AI, an AI-driven mule-account and money-laundering detection system live across 30+ Indian banks with 80% better performance than incumbent solutions; also collaborated directly with UCO Bank’s EFRMS/AML/MIS to integrate the framework.
Big Data Architecture: Architected unified PySpark pipelines for ETL, feature engineering, training and inference with SHAP-driven feature selection and Optuna tuning, cutting bank onboarding latency by 60% (7–9 months to 3 months) and resource consumption by 40%.
Privacy-Preserving Machine Learning: Formulated and applied PPML frameworks to safely train and evaluate models on inter-bank distributed transaction datasets while strictly meeting banking regulatory constraints.
AI Governance & Drift Detection: Engineered an early four-stage drift detection and Model Risk Management framework for automated retraining and continuous performance validation under dynamic fraud typologies, backed by Power BI governance dashboards.
Decision Engine & Threshold Optimisation: Automated feature selection and threshold optimisation for the MuleHunter project using univariate and multivariate hypothesis testing with cost-sensitive loss functions on Precision@K and PR-AUC, cutting processing load by 90% and balancing false-positive customer friction against fraud loss.
Project Ecosystem (ULI, IDPIC, I4C Trail Hunter): Led technical implementation for the Digital Payments Intelligence Platform (IDPIP) and Trail Hunter, deploying agentic LLM pipelines to automate Suspicious Transaction Report (STR) drafting.

Executive Research (AI & NLP)

Feedsense AI Pvt Ltd · Kolkata, India · Apr 2024 – Jan 2025

Generative AI & RAG Systems: Designed end-to-end Retrieval-Augmented Generation (RAG) pipelines over unstructured financial reports using LangChain, vector databases and knowledge graphs for compliance systems, reducing research analysis turnaround by 70%.
Content Analysis (NLP): Formulated content analysis algorithms to extract entity relationships, financial indicators and compliance metrics from large volumes of unstructured documentation.
ETL & Text Analytics: Built scalable ETL workflows and automated data ingestion architectures for high-volume text analytics over filings, disclosures and news feeds.
MLOps & Model Governance: Established end-to-end MLOps pipelines (MLflow, Git-based CI/CD) automating deployment, statistical drift detection, model monitoring and reproducible retraining, ensuring model accuracy, governance and scalability.

Executive Research (AI & ML)

Vista Intelligence Pvt Ltd · Kolkata, India · Jan 2023 – Mar 2024

Multimodal Deep Learning: Engineered multimodal Time-Series Transformers and CNN-LSTM networks fusing structured market ticks with unstructured news sentiment and audio signals, achieving >75% directional accuracy and 34% XIRR.
Speech Analytics (ASR): Designed and tuned RNN-Transducer speech recognition pipelines across diverse Indian accents, slashing Word Error Rate (WER) from 56.8% to 23.4% for real-time audio analytics.
Predictive Modelling & Validation: Validated predictive modelling architectures with time-based validation splits, out-of-sample stress testing and walk-forward backtesting, ensuring high performance and accuracy.
Large-Scale Data Engineering: Programmed robust Python and SQL ETL pipelines for market data processing, translating statistical outputs into strategic investment insights for non-technical stakeholders.

Education

MSc Big Data Analytics (Data Science)

RKMVERI · Belur, West Bengal, India · 2021 – 2023

Relevant courses: Probability & Stochastic Process, Statistics & Statistical Inference, Machine Learning, Deep Learning, NLP, Computer Vision, Optimization Techniques, Linear Algebra, Time Series Analysis, Survival Analysis, Multivariate Statistical Analysis, Cloud Computing, Data Mining, Data Structures & Algorithms, DBMS.

BSc Mathematics

Vidyasagar University · Medinipur, West Bengal, India · 2017 – 2020

Relevant courses: Set Theory, Calculus, Higher Algebra, Real Analysis, Differential Equations & Vector Calculus, Group and Ring Theory, Graph Theory, Probability & Statistics, Numerical Methods, Operation Research, DBMS.

Skills

Primary & Core Technologies: Python, R, Apache Spark (PySpark, SparkML), Kafka, SQL, Hive.
Deep Learning & NLP: Neural Networks (CNN, LSTM, Transformers, GNN, GANs, VAEs), Generative AI, LLMs, RAG (LangChain, LlamaIndex, Ollama), Vector Databases, Knowledge Graphs, Entity Relationship Extraction, Topic Modeling, PEFT (LoRA, QLoRA), Text & Speech Analytics (RNN-T, ASR).
Fraud, Banking & Governance: AML/CFT, mule-account detection, Transaction Risk Scoring, Anomaly Detection, Privacy-Preserving ML, AI Governance & Lifecycle Management, Drift Detection, Model Risk Management (MRM), Threshold & Cost-Benefit Optimisation, Class-Imbalance Modelling.
Machine Learning & Frameworks: XGBoost, LightGBM, Random Forest, Scikit-learn, PyTorch, TensorFlow, Keras, Optuna, H2O, SHAP, SpaCy, OpenCV, Feature Engineering, Hyper-parameter Tuning, Transfer Learning.
MLOps & Big Data Tools: Docker, MLflow, Airflow, CI/CD, DuckDB, Git, Azure AI Studio, Azure Synapse, Azure SQL DB, Azure Data Explorer.
Statistics & Data Analysis: Statistical Inference, Hypothesis Testing, Bayesian Methods, Time Series Analysis, Survival Analysis, Model Evaluation (Precision@K, PR-AUC, ROC-AUC, KS, PSI), Data Mining, Web Scraping.
Visualization & Reporting: Microsoft Power BI, Tableau, Streamlit, Plotly, Seaborn, Matplotlib, Excel.
Soft Skills: Problem-Solving, Stakeholder Management, Cross-Functional Collaboration, Communication, Analytical Thinking, Adaptability.

Research & Publication

Investigate How Market Behaves: Toward an Explanatory Multitasking Based Analytical Model for Financial Investments — IEEE Access, March 2024. More details

Courses & Certifications

Azure Data Scientist Associate (June 2024) — Microsoft
Artificial Intelligence (AI) for Investments (April 2023) — NPTEL
Cloud Computing and Distributed Systems (March 2023) — NPTEL
NISM-Series-XV: Research Analyst (Feb 2023) — National Institute of Securities Markets
Data Base Management System (Oct 2022) — NPTEL
Deep Learning for Computer Vision (Oct 2022) — NPTEL
Data Science Math Skills (April 2020) — Duke University, Coursera

All certificates with details

Personal Projects

Fin-Bot: Advanced Agent-based Financial Chatbot (demo video) — NLP, LLM, Generative AI, RAG, Agentic Workflows
- Employed an LLM-equipped agent to route user queries to the relevant web or custom database, ensuring up-to-date, comprehensive insights.
- Implemented a ChromaDB-based vector database for efficient retrieval of financial news articles and regulatory reports.
- Seamlessly integrated web search functionality ensuring comprehensive responses to compliance and research queries.
Deep Bidirectional LSTM Network for Textual Sentiment Analysis (GitHub) — Deep Learning, NLP, Sentiment Analysis, Web Scraping
- Integrated the Twitter API for real-time tweet scraping and streaming ingestion.
- Leveraged a Bi-LSTM architecture to process textual data for sentiment classification.

More projects are on the Projects page.