import streamlit as st st.markdown( """ """, unsafe_allow_html=True, ) st.html( """ Debopam Chowdhury (Param)
Debopam Chowdhury (Param)

Debopam Chowdhury (Param)

Creator of AiGymBuddy.in | Machine Learning | Deep Learning | Flutter | Math | MLops | TensorFlow | FastAPI

(He/Him)

About Me

Passionate about Machine Learning, Computer Science and Mathematics. Has a strong grip over Machine & Deep Learning Fundamentals, Computer Networks, OS and DSA. Solved 100+ problems on leetcode. Have a keen interest in learning Mathematics, Deep Learning and Statistics. I like to understand things in a deeper way.

Skills

Technical

  • Programming: Python, Java, Dart, VS Code, Git, GitHub, Jupyter Notebooks, CI/CD
  • Machine Learning & Deep Learning: TensorFlow, PyTorch, Scikit-learn, Keras, Supervised Learning, Unsupervised Learning, Neural Networks, Sequence Modeling, Convolution, Attention Mechanisms, Transformer, GPT, BERT, Hyperparameter Optimization
  • Data Handling: Pandas, NumPy, Data Manipulation, Data Preparation, SQL, Pyspark, NoSQL
  • Cloud & DevOps: Docker, AWS, Cloud-AI, FastAPI
  • Mathematics: Linear Algebra, Probability, Statistics, Boosting Methods
  • Frameworks & Tools: Flutter, Firebase, Deep Learning Frameworks, Model Training & Optimization, Version Control, LLM FineTuning
  • Generative AI: Vector Embeddings, Indexing, Chunking, RAG pipelines, LlamaIndex, LangChain, Colpali, Byaldi, Chroma DB

Experience

GENERATIVE AI ENGINEER (Contract - Remote)

Private Client | Sydney, Australia

October - November 2024

Projects

LLM-finetuning and SQL Agent with Auto_Execution with DuckDB, Schema Retriever from CSVs, manual SQL executer.

Current WebAPP which you are using

Deployed

Deep Learning Based Recommendation System

Initially, I aimed to build a recommendation system from scratch using TensorFlow Recommenders (TFRS) on the MovieLens 1M dataset. This involved creating user and movie embeddings with a candidate generation and ranking model.

However, this TFRS approach proved too resource-intensive and time-consuming for effective training and testing. Crucially, the initial results weren't satisfactory for deployment.

Therefore, for the deployed web application, I switched to pre-trained models (BGE embeddings and re-ranking). This offered:

  • Better Performance: More relevant recommendations.
  • Reduced Resources/Time: Faster training and deployment.

While the TFRS code is also included in the web app, the pre-trained model approach was chosen for its superior results and efficiency in a deployment setting. A future improvement could be fine-tuning the pre-trained models for even better performance.

Technologies used: TensorFlow Recommenders, Scann, Vector DB, Distributed GPU Training, Langchain, Streamlit, BAAI BGE Models

Try it out APP Live

Deployed

IBM EMPLOYEE ATTRITION PREDICTOR (End to End with Deployment to AWS, FastAPI with Proxy Server)

Objective: Predicted employee attrition with 85% AUC to improve employee retention and business performance.

Model Development: Hyperparameter optimized Multi-Layer Perceptron, XGBoost, Logistic Regression with Inference as well as Training Pipeline

Backend: Developed a FastAPI backend for real-time predictions, using Pydantic for schema validation for incoming and outgoing requests

Deployment: Containerized with Docker, deployed on AWS EC2, managed via AWS ECR.

CI/CD: Set up an automated CI/CD pipeline using GitHub Actions for seamless updates.

Web Application: Built a user-friendly interface using Flutter Web for real-time interaction.

Security: Handled HTTPS requests using Caddy as a reverse proxy server

Technologies used: TensorFlow, AWS, Docker, FastAPI, CI/CD Pipeline, Multi-Layer Perceptron, Neural Network, XGBoost, Logistic Regression, Hyperparameter Tuned Models, GitHub Actions, Pydantic, Flutter Web, Reverse-Proxy-Server: Caddy

Explanation & Live Demo Click to try out Live

Deployed

AI GYM BUDDY (Langchain | Flutter | Riverpod | Gemini)

Personalized AI-Driven Workouts with Smart Equipment Detection and Progress Tracking

Features: Al Instrument Detection (Camera or Gallery), Exercises based on Available Equipments, Time, Preffered Muscle Groups & Custom requests, Dynamic Video Tutorial Finder for each exercise, Super personalized Al generated routine, Workout History Tracker, Easy SignUp/Login with Google Oauth

Technologies used: Dart, flutter, firebase, gemini 1.5 flash, riverpod, langchain, fastapi, google oauth

Licenses: This code of this app/website is written from scratch and I hold all the rights over distribution

1 Minute Video Demo Google Play Store Google Play Store (Android) Prototype for WEB www.aigymbuddy.in

Deployed

Non-Sequential Breast Cancer Classification System

Multi-Modal Cancer Detection: Developed a novel multi-output deep learning model for breast cancer detection, predicting cancer presence, invasiveness, and difficult-negative case status. The model incorporates both mammogram images and tabular clinical data, leveraging a non-sequential architecture to process distinct data modalities.

Fine-Tuned Image Feature Extraction: Utilized a pre-trained EfficientNetV2B3 model for image feature extraction, fine-tuning layers from block 6 onwards to enhance its applicability to the specific task, thus improving the quality of learned representations and potentially making the model more robust and accurate.

Distributed Training: Accelerated model training through distributed training using TensorFlow's MirroredStrategy on 2xT4 GPUs for 9 hours on Kaggle, demonstrating proficiency in optimizing model training with limited computational resources.

Technologies used: TensorFlow, Transfer Learning, EfficientNetV2, Fused MB-CNN

Live Webapp + Architecture + Training Code + Evaluation Metrics

Deployed

Image Entity Extraction with Qwen2 VL: Large-Scale Inference

Problem Statement: E-commerce and healthcare industries struggle to efficiently extract product details (weight, volume, dimensions) from images at scale.

Action: Developed a large-scale image-to-text inference pipeline using Qwen2 VL: 2B, incorporating image preprocessing, Regex, and parallel processing. Processed 84,000 of 131,000 test images.

Result: Successfully extracted product values from a significant portion of the dataset. Our team of four ranked 172nd out of ~75,000 in the Amazon ML Challenge with Fl-Score=0.47, demonstrating the solution's potential for automated product information extraction.

Technologies used: Qwen2 VL, Python, Regex, Parallel Processing

Click Here to see the code

LLM based ATS System using VertexAI Embedding

Technologies used: Langchain, VertexAI Embedding, StreamLit, PostGresVector

Ongoing

Volunteering

Git/GitHub Instructor (Volunteer)

Carried out sessions to teach juniors the fundamentals of Git and GitHub, covering version control, collaboration, and best practices.

Event Coordinator

Acharya Technical Club - Steigen

Achievements

Certifications

Education

BE in Information Science

Acharya Institute of Technology, Bangalore

2021-2025

CGPA-8.12

Higher Secondary Education

Kalyani Public School, Barasat, Kolkata

2021

77% (Auto Pass Covid Batch)

Secondary Education

Sacred Heart Day High School, Kolkata

2019

90%

""" )