• 🧩 True 4-bit Neural Network Architectures
  Trained SimpleResNet4bit & VGG4bit from scratch using STE on CPU, achieving 86%+ accuracy on CIFAR-10 with <1MB memory.

• 📱 Offline RAG for Mobile & Embedded
  Fully offline GGUF-powered RAG app (React Native) with native integration of llama.cpp + local vector store.

• 🧪 Open Researcher
  I openly share models, experiments, and architectures via arXiv, Zenodo, and GitHub.

• 🛠 Tool-Building Agents
  Researching autonomous agent architectures that create tools using LangChain, CrewAI, and low-level orchestration.

• 🔬 True 4-bit Quantized Networks
  → World's smallest VGG achieving 88.43% on CIFAR-10 using symmetric quant + STE — trained on dual-core CPU.

• 📱 DevShakti Offline RAG App
  → React Native + GGUF + vector search, fully offline chatbot, 100% on-device LLM inferencing.

• 🔗 AI Agents for Tool Creation
  → An open-source project building LangChain/CrewAI-based agents that build their own Python tools.

• 📄 arXiv Draft: "True 4-bit Quantization of Deep Neural Networks Trained from Scratch"
• 💾 Zenodo: https://zenodo.org/record/1234567 (coming soon)
• 🧪 Releasing all experiments: training logs, inference demos, and visualization notebooks.

• 📧 Email: [email protected]
• 🐦 Twitter: @shiv_tathe
• 💼 LinkedIn: Shivnath Tathe
• ☕ Support: Buy Me a Coffee

"I trained a neural network on a dual-core CPU while the world chased TPUs."

© 2025 Shivnath Tathe. All Rights Reserved.