arXiv • Zenodo • Google Scholar • LinkedIn • HuggingFace
I'm Shivnath Tathe, a Software Engineer at ISG eSolutions and an independent AI researcher from India. I work on training neural networks at extremely low precision, proving that 4-bit quantized models can match full-precision accuracy without expensive GPUs.
My first paper on arXiv demonstrates training a convolutional network from scratch at true 4-bit precision on a standard CPU, achieving 92.34% on CIFAR-10 with 8x memory compression. I'm currently building T4NT, a 1.5B parameter multilingual Indian language model trained from scratch on 10 languages using 4-bit quantization-aware training with tanh soft clipping.
I believe powerful AI should not require powerful hardware.
| Paper | Venue | Links |
|---|---|---|
| True 4-Bit Quantized Convolutional Neural Network Training on CPU: Achieving Full-Precision Parity | arXiv (cs.LG) | Paper · Code |
| Autonomous Tool-Creation in AI Agents: A Conceptual Framework for Self-Evolving Systems | Zenodo | Paper · Code |
4-bit Quantization-Aware Training
- Trained VGG-style networks at true 4-bit precision from scratch using symmetric quantization + straight-through estimators
- 92.34% on CIFAR-10 (0.16% gap from full-precision) with 8x memory compression
- Validated on CIFAR-100 (70.94%) and on mobile (OnePlus 9R, 83.16% in 6 epochs)
- No specialized GPU kernels. Standard PyTorch on CPU.
T4NT-1.5B (in progress)
- Multilingual Indian LLM trained from scratch on 10 languages
- Architecture: RMSNorm + RoPE + SwiGLU + 4-bit QAT + Tanh Soft Clipping
- Custom SentencePiece tokenizer (65K vocab) covering Devanagari, Bengali, Tamil, Telugu, Kannada, Malayalam, Gurmukhi scripts
- Training on Kaggle TPU v5e-8
| Project | Description |
|---|---|
| true-4bit-training | 4-bit QAT with tanh soft clipping — arXiv published |
| DevShakti Offline RAG | React Native + GGUF, fully offline on-device LLM chatbot |
| AgentForge | LangChain/CrewAI agents that build their own tools |
Research : PyTorch, Quantization, QAT, STE, LoRA, PEFT
Models : llama.cpp, GGUF, HuggingFace Transformers
Frontend : React Native, Angular, Electron
Backend : FastAPI, Node.js
Infra : Kaggle TPU, Google Colab, Linux