About

Hi! I am Aditya. I am a Data Scientist with 4 years of experience spanning Machine Learning, and Generative & Agentic AI. At Accenture, I build production-grade GenAI systems — RAG-based knowledge assistants and Agentic AI chatbots — for a leading Indian public-sector bank, and have delivered Agentic AI POCs for banks in Southeast Asia. My roots are in Statistics, and I love turning complex data problems into deployed, measurable solutions.

The Journey:

My journey began with a B.Sc. (Hons.) in Statistics from the University of Calcutta and an M.Sc. in Data Science from the University of Kalyani. At Cloudcraftz Solutions in Kolkata, I worked across fraud detection, cashflow forecasting, explainable AI, and NLP pipelines — learning to bridge technical depth with actionable business insight. Since 2024, I have been at Accenture as a Decision Science consultant (promoted from Analyst to Consultant in June 2026), working on-site with a large Indian public-sector banking client. My focus has shifted to Generative and Agentic AI: architecting Azure-based RAG solutions, orchestrating multi-agent systems with LangGraph, modernizing legacy ML estates, and benchmarking enterprise AI maturity. I enjoy owning solutions end to end — from architecture and effort estimation to security hardening and monitoring.

Recent Work:

- Designing and building RAG-based knowledge assistants and Agentic AI chatbots for a major Indian public-sector bank on Azure (Azure OpenAI GPT-4o, AI Search, Document Intelligence, Cosmos DB), orchestrated with LangChain/LangGraph and served via FastAPI — including end-to-end solution architecture, effort estimation, and VAPT security remediation (JWT validation, CORS, authenticated document access).
- Built a POC Agentic AI Personal Finance Advisory agent for an Indonesian bank using a client-hosted Llama model, LangGraph, FAISS, and FastAPI, covering investment profiling, personalized advice, investment education, and lead generation.
- Modernizing the bank's legacy ML estate: migrating R models to Python with XGBoost + ARIMA ensembles, building Hive/Parquet pipelines on Cloudera CML, and authoring an Evidently-based model monitoring pipeline.
- Contributing to enterprise AI maturity audits — benchmarking the bank's personalization/Next-Best-Action capabilities across 13 propensity models spanning Retail, Liability, MSME, Wealth, and Collections.


Explore some of my notable work and projects in the links below!

Floating Image

Contact

Skills

AI/ML Domains

  • Machine Learning
  • Deep Learning
  • Natural Language Processing (NLP)
  • Prompt Engineering
  • Agentic AI & Multi-Agent Systems
  • Generative AI
  • Retrieval-Augmented Generation (RAG)
  • Multi-Agent Orchestration
  • Exploratory Data Analysis (EDA)
  • Statistical Modelling
  • Time Series Analysis and Forecasting

Language and Libraries

  • Python
  • NumPy / Pandas
  • Scikit-Learn / XGBoost
  • PyTorch / TensorFlow
  • HuggingFace
  • FastAPI / Streamlit
  • LangChain / LangGraph
  • Azure OpenAI & AI Search
  • Cosmos DB / FAISS
  • Hive SQL / Cloudera CML
  • Evidently (Model Monitoring)

Soft

  • Adaptibility
  • Problem Solving
  • Customer Service
  • Communication Skills
  • Time Mangement
  • Critical Thinking
  • Collaborative Work
  • Presentation Skills

Experience

  • Ind & Func Decision Science Consultant
    Accenture, Mumbai

    July 2024 - Present  ·  Promoted from Analyst to Consultant in June 2026
    • GenAI Knowledge Assistants & Agentic AI Chatbots for a Leading Indian Public-Sector Bank

      • Building and maintaining RAG-based knowledge assistants and Agentic AI chatbots using Azure OpenAI (GPT-4o), Azure AI Search, Cosmos DB, LangChain/LangGraph, and FastAPI.

      • Designed the full-scale Azure RAG architecture end to end: Function App ingestion, Document Intelligence, embeddings, vector search, Content Safety, Key Vault, authentication, Cosmos DB logging, and LLM evaluation — along with the admin portal and chatbot frontend.

      • Led VAPT remediation for the production chatbot, hardening the backend with JWT validation (audience, issuer, signature, expiry), CORS fixes, and authenticated document access.

      • Prepared component-wise effort estimation and commercial ballparks for the RAG solution, translating architecture into person-day and pricing models.

    • Agentic AI Personal Finance Advisory POC — Indonesian Bank

      • Built a conversational Agentic AI advisor hosted on the bank's application, using a client-hosted Llama API, LangGraph, FastAPI, and FAISS, with a Streamlit interface visualizing agent interactions.

      • Delivered investment-profile creation, personalized portfolio-based advice, investment education, and lead generation by detecting customer interest in financial products.

    • Legacy ML Modernization (R → Python)

      • Migrating the bank's legacy R models (including a Term Loan Prepayment rate model) to Python using XGBoost + ARIMA ensembles.

      • Built Hive/Parquet data pipelines on Cloudera CML and HDFS, and authored an Evidently-based model monitoring pipeline.

    • Enterprise AI Maturity Audit & Benchmarking

      • Assessing the bank's Next-Best-Action/personalization engine maturity across 13 propensity models spanning Retail, Liability, MSME, Wealth, and Collections.

    • GenAI POC Support — Second Indonesian Bank

      • Contributed prompt engineering for a Generative AI proof of concept.

  • Data Scientist
    Cloudcraftz Solutions

    July 2022-July 2024  ·  Promoted from Analyst to Consultant in June 2026
    • Commodity Price Change Prediction

      • Developed an advanced multi-classifier pipeline that accurately predicts price fluctuations (in percentage) of commodities within a specific time-frame.

      • Spearheaded the creation and testing of prompts on cutting-edge language models (GPT, LLaMa, Gemini) using LangChain, enabling the extraction of sentiment and relevance scores from news articles to forecast commodity prices.

      • Leading a pioneering collaboration to integrate textual data with the multi-classifier pipeline, revolutionizing business applications in the industry.

      High Frequency Trading Strategy Analysis

      • Conducted thorough evaluations and analyses of various trading strategies, leveraging data visualization techniques and statistical tests to optimize performance.

      • Delivered actionable insights for enhancing trading strategies, driving profitability and success in high-frequency trading environments.

      Strategic Development of Transaction Network Visualization and Graph Neural Network Models for Enhanced Financial Security

      • Developed innovative network visualization solutions for improved understanding of complex financial data.

      • Spearheaded the identification of suspicious nodes through comprehensive analyses:

        1. Analyzed transaction volume, centrality metrics, and transaction amounts.

        2. Deployed advanced techniques to pinpoint anomalies and potential fraud indicators.

      • Pioneering advancements in outlier detection:

        1. Actively learning and utilizing PyTorch Geometry for Graph Neural Network development.

        2. Aiming to refine outlier detection models for robust network security and enhanced fraud prevention.

    • Enhanced Financial Decision-Making with Innovative Cashflow Forecasting Method

      • Collaborated on developing a sophisticated cashflow forecasting tool for an NBFC client to enhance financial decision-making.

      • Utilized PgAdmin to preprocess raw financial data, translating complex insights into intuitive visualizations.

      • Conducted extensive feature engineering to improve predictive accuracy, resulting in a robust machine learning model.

      • Achieved a remarkable 2.9% reduction in RMSE through a stacked model approach, showcasing significant improvement in forecasting precision.

    • An Explainable AI Product: Innovative Development of Platform to interpret Machine Learning Models with Shapley Values and User-Centric Design

      • Applied Shapley values to interpret machine learning models, providing both global and local explanations for feature contributions and their impact on predictions.

      • Developed visualization and explanation tools for quantifying feature contributions, enhancing model interpretability.

      • Integrated counterfactual-based explanations for user-driven exploration of hypothetical scenarios, improving overall model understanding.

      • Collaborated with cross-functional teams to ensure actionable insights for non-technical stakeholders.

      • Emphasized user-centric design by implementing intuitive interfaces and interactive visualizations to enhance user engagement and comprehension of model output.

    • Empowered Stock Price Predictions with Advanced Sentiment Analysis Pipeline

      • Developed a web scraping and NLP pipeline for efficient gathering and extraction of textual data from diverse news websites.

      • Utilized Hugging Face Transformers for fine-tuning a sentiment analysis model.

      • Achieved high-precision sentiment classification, enhancing the accuracy of stock price predictions.

    • An EDA Platform: Developed User-Friendly EDA Platform for Comprehensive Data Exploration

      • Designed and deployed an in-house Exploratory Data Analysis (EDA) platform for tabular and time series data.

      • Implemented user-friendly visualization tools, enabling non-technical users to make data-driven decisions.

      • Enhanced the platform's statistical analysis capabilities for comprehensive data exploration.

    • Optimized Database Labeling with NLP-Based Services

      • Delivered NLP-based dataset labeling services for an international client.

      • Streamlined the database labeling process through efficient implementation.


  • Research Intern
    USAID Project under LISA 2020, in association with Department of Statistics, University of Calcutta and National Institute of Wind Energy, Government of India

    March, 2022 - July, 2022
    • Optimized Predictive Forecasting with Advanced Regression Time-Series Models

      • Conducted thorough data exploration using rigorous Data Visualization and Exploratory Data Analysis techniques.

      • Implemented advanced Regression-based-Time-Series Models to enhance predictive forecasting of the GHI.

      • Achieved an impressive R-squared score of 0.92, showcasing the model's high predictive accuracy.


  • Research Intern
    A. K. Choudhury School of IT, University of Calcutta

    September, 2021 - July, 2022
    • Environmental Sound Classification with CNN Models

      • Engaged in a hands-on project on Environmental Sound Classification using the ESC-50 dataset.

      • Applied audio processing techniques to modify and extract essential spectrograms.

      • Implemented Convolutional Neural Network (CNN) models for sound classification tasks.

      • Demonstrated consistent proficiency with an impressive average accuracy score of 87%.

      • Download Resumé
      Download Cover Letter

Personal Project

  • AI Culinary Symphony: Crafting Efficient Large Language Models with Precision and Flavor

    • Description

        Designed and implemented an end-to-end training pipeline for Large Language Models (LLMs), akin to crafting a gourmet meal. This involved meticulous data gathering, diverse prompts creation, and leveraging key libraries (trl, peft, transformers, torch) for effective fine-tuning. The process incorporated optimization techniques (wandb, einops, pandas, datasets, accelerate, bitsandbytes) for efficient model training.

    • Processes Showcased

      • Monitored GPU usage and provided concise training progress reports, ensuring efficient model learning without resource bottlenecks.

      • Ensured reproducibility across experiments, guaranteeing consistent and predictable results.

      • Efficiently gathered and prepared training and validation datasets, optimizing the model's data ingestion process.

      • Crafted a customizable tokenizer for precise language dissection, allowing for tailored learning experiences.

      • Implemented quantization and LORA techniques for model size reduction and optimization.

      • Constructed an intelligent and efficient language model, considering PEFT-enabled architectures and k-bit training for resource-conscious brilliance.

    • Outcome

        Successfully orchestrated a comprehensive training pipeline, balancing flavor-rich language model development with resource efficiency. Achieved a model capable of sophisticated language tasks while respecting computational constraints.

    • Tech-stack
        PyTorch, Hugging Face Hub, WandB, trl, peft, transformers
    • Notebooks and Spaces