I train neural networks at extreme efficiency.
Shivnath Tathe is a Software Engineer and independent researcher specializing in 4-bit neural networks, RAG systems, and autonomous AI agents. Based in Hyderabad, India.
Shivnath Tathe is a Software Engineer and independent researcher based in Hyderabad, India, specializing in 4-bit neural network training, retrieval-augmented generation systems, and autonomous AI agent design. He works as a Software Engineer at ISG eSolutions, where he builds production multi-agent pipelines, RAG applications, and scalable Angular front-ends.
He graduated from Shivaji University in 2024. His research is published on arXiv and Zenodo, and his open-source models are available on Shivnath Tathe's Hugging Face profile. His code and projects are maintained on Shivnath Tathe's GitHub. See his research work below.
Provider-agnostic memory and skills infrastructure for AI agents, delivered through an OAuth 2.1 MCP server. Sutra works across Claude, GPT, Grok, Copilot, Cursor, Windsurf, Gemini CLI, and more than ten other clients, with a Next.js dashboard, Supabase backend, npm CLI, and 15+ beta users.
Visit Sutra
A two-stage RAG architecture where semantic search retrieves chunk identifiers and precomputed KV caches are loaded from NVMe directly into attention. This avoids generation loss from chunked context and reached 60% accuracy on Apple 10-K question answering, compared with 50% for the baseline.
A production enterprise AI platform running Qwen3.6:35B on Apple M4 Pro hardware through Ollama, LiteLLM, Open WebUI, Cloudflare tunnels, and Tailscale. It powers LangGraph-based email automation with RabbitMQ, Brevo, and SQL Server CRM integration, plus PulseIQ morning briefings for the full team.
A local MCP server for live communication between AI agents from different vendors. OpenBridge demonstrated GPT-6.1-Sol orchestrating GPT-5.6-Sol and Claude Sonnet 5.5 in a relay loop, with opencode HTTP forwarding, Claude Code MCP Channels support, and SHA-256 session logs.
View OpenBridge on npm
Converts authorized software into verified MCP tools without LLM-generated code. REA evidence and recorded workflows feed a rule-based compiler that moves structured capabilities through a DISCOVERED, COMPILED, VALIDATED, and ENABLED lifecycle. The MVP passes all three capability tests.
A hands-free coding environment built from opencode with local Kokoro TTS through WebGPU/WASM, whisper.cpp through native Node.js bindings, and FFmpeg. It requires no cloud speech API and is published to npm as voxcode-ai.
View VoxCode on GitHub
A covert messaging system that tunnels communication over UDP port 53 using DNS wire format, designed to remain available during internet shutdowns. It combines X25519 key exchange, AES-256-GCM encryption, ciphertext-only server storage, a React Native mobile app, and an AWS EC2 relay.
A local transcription pipeline that attaches to Teams through CDP, captures live captions and speaker names from the rendered DOM, and combines them with WASAPI audio and faster-whisper for multilingual English, Hindi, and Marathi meetings. It produces dated transcripts and daily AI-assisted summaries.
A developer utility that reads opencode's SQLite database, detects project anchors pointing to missing directories, and applies controlled path remapping through a local web interface. It has been tested against 305 live anchors.
A scheduled daily email digest that selects ten technology stories at 07:00 IST using Gemini with a Grok fallback. The production service uses FastAPI, APScheduler, Supabase, Gmail SMTP, Render, and a Cloudflare-managed domain.
Visit TechPulse Daily
A production OCR platform built around the 3B MoE Baidu Unlimited-OCR model. It serves inference from a private VPS through llama-server, LiteLLM, FastAPI authentication and rate-limiting middleware, a Cloudflare tunnel, and a React frontend.
A drop-in reverse proxy for the NVIDIA NIM free tier, using an AsyncIO priority queue and sliding-window counter to absorb rate limits without requiring client changes. Stress testing completed with zero leaked HTTP 429 responses.
View UnlimitedNIM on GitHub
A 0.5B-parameter decoder pretrained on FineWeb-Edu with JAX on a Kaggle TPU v5e-8 and the Qwen2.5 tokenizer. The dataset is published on Hugging Face, with a domain-specialized Shakti-SQLCoder-0.5B model as the target.
An experimental attention architecture where prefix tokens use a low-cost linear projection and broadcast global context instead of quadratic self-attention. Tests on a 124M-parameter model identified content-relational orthogonality as the central obstacle to retrofitting pretrained models.
Research into training language models so weight vectors cluster by domain, allowing input-directed sparse loading from SSD. Experiments use Llama-3.2-1B, with a pairwise kernel loss identified as the appropriate training objective and three supporting theorems drafted.
A competition system for abstract visual reasoning on ARC-AGI-3. The 25-game baseline reached a mean RHAE of 0.89 on an RTX Pro 6000, using parallel solver diversity and consensus-based answer selection.
An Ollama-style Python CLI and patched llama.cpp backend that keeps the KV cache on SSD through memory mapping instead of allocating it entirely in private RAM. It achieved a maximum measured 57x RAM reduction at 256K context and 2.7 tokens per second with a 35B MoE model on a 16GB laptop. The architecture overlaps Apple's "LLM in a Flash" approach and exposes an OpenAI-compatible local API.
View DiskLLM on GitHub
A local-first LLM observability SDK with 3,600+ PyPI downloads. It instruments OpenAI, Anthropic, LiteLLM, Qdrant, ChromaDB, and Pinecone through OpenTelemetry, while keeping traces in a single-file SQLite backend with no cloud dependency.
View OpenSmith on GitHub
An educational vector database implemented in approximately 130 lines of NumPy-based code. It builds co-occurrence word embeddings, applies SVD dimensionality reduction, averages vectors into document embeddings, and performs brute-force cosine search while making each stage of the retrieval pipeline readable.
View nanoVectorDB on GitHub
This paper demonstrates that CNNs can be trained from scratch at true 4-bit precision
on commodity CPUs, with no specialized hardware and no post-training tricks. It achieved
92.34% accuracy on CIFAR-10 (just 0.16% below full-precision baseline), 70.94% on CIFAR-100,
and 83.16% on a consumer Android device in only 6 epochs, with 8x memory compression over FP32.
The key contribution is tanh-based soft weight clipping as a novel quantization technique.
Read the paper on arXiv
The same 535M parameter decoder-only architecture as T4NT-0.5B, extended with training on
50M tokens of WikiText data. Uses tanh soft weight clipping throughout and achieves improved
perplexity over the 15M-token baseline. Released under Apache 2.0.
View
T4NT-0.5B-50M on Hugging Face
A 535M parameter decoder-only transformer trained entirely from scratch using 4-bit
quantization-aware training on a single NVIDIA T4 GPU in under 3 hours. Architecture
uses RoPE, SwiGLU, and RMSNorm. Best validation perplexity of 93.04 on WikiText-103.
Deployed size of only 255 MB in INT4. Over 200 downloads on Hugging Face.
View T4NT-0.5B
on Hugging Face
Instruction-tuned version of T4NT-0.5B, fine-tuned to follow user prompts and
instructions rather than purely completing text. Retains the 4-bit quantized weight
structure with full instruction-following capability.
View
on Hugging Face
The working implementation of the Zenodo paper on autonomous tool-creation.
Agents are built with LangChain and CrewAI and dynamically generate new tools at runtime
to handle unforeseen challenges, reducing human intervention in complex workflows.
View AgentForge on GitHub
A conceptual framework for AI agents capable of self-generating tools in real-time.
Targets mission-critical environments including aerospace, medical emergencies, and defense.
Explores theoretical foundations, potential applications, and open challenges in building
self-evolving agentic systems that require no human intervention.
Read the paper on
Zenodo
A self-hosted OpenAI-compatible platform running a QLoRA fine-tuned Qwen2.5-1.5B model
trained on India-focused synthetic data. It includes a FastAPI backend, the devshakti-ai
Python SDK, and a React playground for testing local inference.
Visit DevShakti AI
A full-stack AI content generation platform capable of producing 12+ content types including blog titles, SEO keywords, and social media captions. Built with Next.js for SSR, Drizzle ORM with PostgreSQL for data handling, and Razorpay for credit-based payments. Achieved 30% faster page loads through backend optimisation.
DevShakti is a fully offline Android app built with React Native, running GGUF-quantized
LLMs entirely on-device with no cloud dependency and no data leaving the device. Features
include document upload with on-device RAG, tool calling, multi-agent pipelines, and the
ability to connect external apps directly into the assistant.
View DevShakti Android releases on GitHub
Building RAG-based knowledge base systems using FastAPI, OpenAI, and vector embeddings, serving 10+ business clients through a 3-stage retrieval pipeline. Migrated vector storage from Pinecone to Qdrant for cost efficiency, adopted by 20+ clients in production. Built AI workflow automation tools including an offline LLM SQL assistant and a travel itinerary generator that cut extraction time by 57%. Designing custom multi-agent pipelines using LangGraph and LangChain for production workflows.
Computer science and software engineering.
Zenodo preprint · July 2026
INTERN lets a language model autonomously acquire domain knowledge before deployment by exploring business systems, extracting facts, and using them for adapter-based fine-tuning. A configurable human-in-the-loop gate supports environments with lower data confidence.
The position paper defines exploration and operating phases plus deferred and incremental adapter-merge strategies, without architectural changes or labeled training data.
arXiv:2603.28611 · March 2026
LACE introduces a loss-driven mechanism that dynamically expands neural network capacity during training. When sustained loss indicates insufficient capacity, new dimensions are added without labels, replay buffers, or external control.
Results: Precise domain-boundary detection (100% precision), performance matching large fixed-capacity models, and unsupervised domain separation in GPT-2 activations.
arXiv:2603.13931 · March 2026
This paper challenges the assumption that neural network training requires high precision and expensive hardware. Using a novel tanh-based soft weight clipping technique combined with symmetric 4-bit quantization, dynamic per-layer scaling, and straight-through estimators, a VGG-style CNN was trained entirely from scratch on commodity CPUs with no specialized kernels and no post-training quantization.
Results: 92.34% accuracy on CIFAR-10 (0.16% below full-precision baseline), 70.94% on CIFAR-100, and 83.16% on a consumer Android device in 6 epochs. 8x memory compression over FP32. Source code, training logs, and model checkpoints are open-source.
Zenodo · 2025
This paper proposes an agent architecture that enables autonomous tool generation during execution, going beyond AutoGPT and LangChain-style agents restricted to predefined toolsets. The framework allows agents to identify gaps in their capabilities and synthesize new tools on the fly without human intervention.
Key applications explored include autonomous diagnostics, emergency-response systems, aerospace, and defense. Implemented as AgentForge, an open-source project built on LangChain and CrewAI.
Email Email Shivnath Tathe
GitHub Shivnath Tathe on GitHub
LinkedIn Shivnath Tathe on LinkedIn
HuggingFace Shivnath Tathe on Hugging Face
Medium Shivnath Tathe on Medium
DevTo Shivnath Tathe on DevTo