Machine Learning & Software Engineer with expertise in ML/DL, MLOps, Mathematics, RAG, and Flutter. I'm passionate about building deployable AI-based solutions and have a strong foundation in Machine Learning, Computer Science, and Mathematics.

• 🌱 Currently working on LLM Pre-Training from scratch using PyTorch & Math
• 💼 Have 10+ Deployed projects from various domains like: NLP/LLM - Computer Vision - Deep Learning - Synthetic Data Generation - GenAI, where all training code and models are open-sourced.
• 🏋🏻‍♂️Creator of AiGymBuddy - PlayStore / Web Prototype / Video Demo
• 🎓 BE in Information Science from Acharya Institute of Technology (2021-2025)
• 📫 Reach me at: [email protected]
• 🌐 Website - https://debopamparam.github.io/DebopamParam/

Private Client, Sydney, Australia | October-November 2024

• Developed secure, on-premise solutions for complex PDF with images and charts Q&A and knowledge retrieval
• Built data ingestion pipeline for 1000's of documents with automatic task scheduler
• Implemented multimodal RAG pipelines (Byaldi, Colqwen2, Pixtral 12B) optimized for diverse document types (70% accuracy improvement)
• Containerized the application with Docker for deployment flexibility

Technologies: NLP, Vision Embeddings, Local Multimodal RAG, LangChain, Pixtral 12B, Col-Qwen2, Byaldi

Small open-source contribution - Byaldi - 575 ✰

Solo Creator: Design - Code - Deploy - Marketing ---- Deployed✅

Personalized AI-driven workout app with smart equipment detection and progress tracking.

• 650+ registered users
• AI instrument detection (camera or gallery)
• Personalized workout routines based on available equipment
• Dynamic video tutorial finder
• Google OAuth integration

Technologies: Dart, Flutter, Firebase, Gemini 2.0, Riverpod, LangChain, FastAPI, Google OAuth, Deep Learning

📱 Google Play Store | 🌍 Website | 🎥 1-Min Demo Video

Focus-flow, Remote | Nov-Dec 2024

UNFILTR, INC, Bengaluru, India | Jan-June 2023

High-accuracy NSFW content detection system for social media platforms ___Deployed✅

• 96% accuracy & 0.92 F1 score
• Incrementally trained on 130,000 sample images
• Two-phase training with EfficientNetV2-M

Technologies: TensorFlow, Incremental Training, Transfer Learning, ReduceLRonPlateau

🌍 Live Webapp + Architecture + Training Code + Training Data

End-to-end ML application predicting employee attrition with 85% AUC ____Deployed✅

• Hyperparameter optimized models (MLP, XGBoost, Logistic Regression)
• FastAPI backend with Pydantic schema validation
• Containerized with Docker and deployed on AWS EC2
• CI/CD pipeline with GitHub Actions

Technologies: 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 Video | 🌍Live Webapp + Architecture + Training Code + Training Data

Scalable recommendation system capable of handling 25M+ candidates ____Deployed✅

• Hybrid architecture with candidate generation and re-ranking
• Custom 4-tower deep learning model trained from scratch using Tensorflow, using Nvidia's 2xT4 GPUs
• Resistant to cold-start problem

Technologies: TensorFlow, Faiss, Vector DB, Distributed GPU Training, Langchain, BGE, Streamlit

🌍 Live Webapp + Architecture + Training Code + Training Data

Multi-output deep learning model for breast cancer detection ____Deployed✅

• Published in IRJET
• Processes both mammogram images and tabular clinical data
• Fine-tuned EfficientNetV2B3 for feature extraction
• Distributed training with TensorFlow's MirroredStrategy in Nvidia 2xT4-GPUs

Technologies: TensorFlow, Transfer Learning, EfficientNetV2, FusedMB-CNN

📃IRJET Published Paper | 🌍 Live Webapp + Architecture + Training Code + Evaluation Metrics

SQL agent created by finetuning Qwen2.5-3B-Coder-Instruct model ____Deployed✅

• Supervised finetuning with QLora, with High Quality SQL synthetic data, generated from ChatGpt-4o.
• Quantized from BF16 to int4(q4_k_m) for lightweight inference
• Integrated with Ollama and LlamaCpp

Technologies: Supervised Finetuning, Unsloth, LlamaCPP, Docker, DuckDB, Langchain, Huggingface Spaces

🌍 Live Webapp + Architecture + Training Code + Training Data | 🖥️ Run Locally Via Ollama

Ongoing (80% Done)

Teaching Qwen2.5-0.5B to learn Bengali language through CPT and adapting it to English-to-Bengali translation using SFT, all under 400MB memory for edge device translation tasks.

Technologies: Unsloth, CPT, SFT, Edge Deployment, Language Model Optimization

Upcoming -- Currently studying all Mathematical Concepts -- Plan to finish this project within 7 April, 2025.

Building a small LLama-style foundational model from scratch using PyTorch, Flash-Attention, and mathematics for less than $50 using Runpod-GPUs.

Technologies: PyTorch, Flash-Attention, LLM Architecture, Runpod-GPUs

1.3K Synthetic SFT dataset made without using any 3rd party Library ____Deployed✅

• Q&A dataset about TurboML with 1,343 technical questions and detailed answers
• Covers implementation, troubleshooting, architecture design, and performance optimization

🌍 Hugging Face Datasets

Made TurboML Chat Agent by scraping all the Docs + Pypi package with Grounding Links ____Deployed✅

🌍 Check it out live

1. Efficient Parallel Implementation of Forward Prop, K-means using Math & Numpy Broadcasting
2. Machine translation using Encoder-Decoder Architecture using Bi-directional LSTMs and Bahdanau Attention
3. Math behind Gating mechanism of a LSTM & GRU cell
4. Learning LangGraph & CrewAi
5. Real-time Data-Center Anomaly Detection in Streaming Data with TurboML (HST + AdaBoost)

Amazon ML Challenge Hackathon Competition - Ranked 172 out of ~75,000 participants

• Developed large-scale image-to-text inference pipeline using Qwen2 VL: 2B
• Incorporated image preprocessing, Regex, and parallel processing
• F1-Score of 0.47

Click Here to see the code

• Ranked 172 out of 75,000 participants in the Amazon ML Challenge Hackathon 2024
• 2nd place out of 60 in the TechnioD Hackathon
• Finalist in IIT Bombay's Mood Indigo Bengaluru Event
• Open-source contribution to Byaldi - 575 ✰: Fix langchain integration not present in pypi tar & whl-- pyproject.toml

• Understand the Importance of Math in ML with Numpy Broadcasting
• Let's Visualize Optimal Floyd's Cycle Detection Algorithm with a String
• Python Compilerrrr???

• Email: [email protected]
• LinkedIn: Connect with me