Skip to content
View soma2000-lang's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report soma2000-lang

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
soma2000-lang/README.md

Somasree Majumder

ML Engineer — Defense CV · Open Source · Systems from Paper to Production

Building real-time computer vision systems for defense at DeepEdge. Contributor to PyTorch Lightning, Hugging Face, and KerasCV/NLP. LinkedIn LIFT Scholar 2022 (Top 500 technologists globally, 0.5% acceptance).

LinkedIn Portfolio LeetCode


What I'm Working On

Anti-UAV Tracking @ DeepEdge — End-to-end production system deployed at 3 defense installations, processing infrared video in real-time on NVIDIA Jetson edge devices.

  • Sub-pixel tracking ensemble (Lucas-Kanade + TAPNet + CoTracker) extending detection range 1.5km → 5km — contributed to $2M+ contract renewal
  • Hybrid CV pipeline (YOLOv8 + Kalman + MOG2/SVM): 95%+ accuracy, 80% false positive reduction
  • 4.2x inference speedup (250ms → 60ms) via TensorRT quantization + CUDA kernel fusion for 30fps edge deployment
  • 3D world-frame localization: monocular depth + camera calibration + IMU fusion → <0.5m error

Open Source Contributions

Project Stars What I Did
PyTorch Lightning 27K+ PRs →
Hugging Face Transformers 135K+ PRs →
KerasCV / KerasNLP Implemented augmentation layers used by 35K+ engineers
Unify.ai (Ivy) Built 15+ TensorRT OpConverters, 3x faster inference, 40% memory reduction

Selected Projects

From-scratch implementations of key deep learning papers in PyTorch. Not wrappers — actual algorithm reimplementations to understand the math.

vector-db — Custom Vector Database

HNSW graph-based approximate nearest neighbor search, built from scratch in Python with minimal dependencies. No LangChain, no Pinecone — just NumPy and the algorithm.

flight-scanner — RAG Flight Analytics

Natural language queries over structured flight data using DeepSeek-R1 (70B) + LLaMA 3.2 (3B) with RAG architecture, ChromaDB, and FastAPI.

amazon-classification — Large-Scale Product Classification

Fine-tuned ALBERT on 500K Amazon listings across 27 categories with 1:150 class imbalance. 89% F1 — beat BERT by 12% and the production baseline by 10%. Deployed with TensorRT.

childmind-prediction — ChildMind Anxiety Prediction

Gated RNNs analyzing children's physical activity patterns for early detection of problematic internet use. Healthcare ML with real clinical relevance.

food-delivery — Full-Stack Delivery Platform

Django REST + JWT auth. Vendor management, proximity search, payment integration, real-time order tracking, push notifications. Shows I build complete products, not just models.


Competition Results

Event Result
Challang Global Hackathon Top 10 / 100 teams worldwide — disaster response CV (satellite imagery + EfficientNet, 91% IoU)
HackerEarth ML Challenge Rank 18 / 2,500
Cipla Datathon Rank 12 / 1,000+ (Top 1.2%)
Google Code Jam for Women Rank 1,600 / 50,000+

Publications

  • "Privacy-Preserving Cloud Computing" — Homomorphic encryption for secure computation. IEMATICS 2020
  • "ML in Medical Diagnosis" — Applied ML for clinical diagnostics. SPECTRUM 2020

Professional Background

DeepEdge (Defense AI Startup) — ML Engineer, Mar 2025 – Present Deloitte (AI Center of Excellence) — ML Engineer, Jul 2023 – Mar 2025 Unify.ai — ML Engineer (Contract), Feb – May 2023

B.Tech ECE, Institute of Engineering & Management, Kolkata — 8.94/10 CGPA


GitHub Stats GitHub Streak

Pinned Loading

  1. Paper-Implementations-in-Deep-Learning Paper-Implementations-in-Deep-Learning Public

    Python 2

  2. Coding-challenges Coding-challenges Public

    ....

    TypeScript

  3. metrics metrics Public

    Forked from Lightning-AI/torchmetrics

    Machine learning metrics for distributed, scalable PyTorch applications.

    Python 1

  4. Amazon-Product-CLassiication Amazon-Product-CLassiication Public

    Used Albert and Distillbert and en- sembled it to improve accuracy by 10.Used synthetic Pytorch datasets.

    Jupyter Notebook 1

  5. My_vector_database My_vector_database Public

    I propose the development of a lightweight and straightforward vector database implementation utilizing Hierarchical Navigable Small World (HNSW) graphs in Python. This solution will be designed wi…

    Python 1

  6. app app Public

    HTML