Nisharg Nargund
I’m an AI researcher, engineer, and builder exploring the future of intelligent systems. My work spans efficient language models, generative AI, edge AI, and AI-powered products, bridging research with real-world engineering.
I’m a co-author of TernaryLM, presented at the GlobalSouthAI Workshop at NeurIPS 2026, with additional research published across IEEE, ACM, and Springer venues. Beyond research, I’ve built startups, open-source projects, and AI systems, always driven by the idea of turning ambitious concepts into things that actually work.
Professional Experience
Founder & Director — OpenRAG Innovations Pvt. Ltd.
Apr 2024 – 2026
- Founded and led an AI startup focused on building reliable, context-aware Generative AI systems for education and enterprise use cases.
- Built DocDynamo, an AI-powered document intelligence platform using RAG and a proprietary corrective context and verification pipeline to improve response accuracy, reduce hallucinations, and optimize LLM context.
- Designed and deployed end-to-end RAG and agentic AI pipelines, covering document processing, retrieval, context optimization, LLM orchestration, and production deployment.
- Grew the platform to 2,600+ users and 2 B2B clients, while developing AI solutions across education and enterprise domains.
- Led product development across research, engineering, deployment, and go-to-market, taking AI systems from early prototypes to production-ready applications.
Researcher — International Institute of Information Technology (IIIT), Hyderabad, India
May 2025 – July 2025
- Implemented BitNet (1-bit LLM) from scratch; evaluated performance under resource-constrained inference on T4 and H100 GPUs.
- Conducted an in-depth architectural analysis of DeepSeek-V3, studying its MoE design, expert routing, and training efficiency innovations.
- Explored edge computing deployment strategies for low-precision LLMs, informing subsequent quantization research (TernaryLM, GlobalSouthAI Workshop @ NeurIPS 2026).
AI Strategy Associate — Strategy Boolean(Remote)
Aug 2025 – Sept 2025
- Built scalable GenAI applications including custom RAG chatbots and LLM pipelines for global clients.
- Optimized prompt engineering and model outputs to reduce hallucinations and improve multilingual support.
- Contributed to production-grade solutions across education and enterprise domains.
Projects
TernaryLM
NeurIPS 2026 · GlobalSouthAI Workshop · AcceptedNative ternary-weight language modeling, inspired by the BitNet paradigm
- Designed and implemented a language model architecture using native ternary weights {-1, 0, +1}, inspired by the BitNet paradigm.
- Built custom BitLinear layers, transformer blocks, training infrastructure, and quantization-aware components rather than simply applying post-training quantization.
- Developed experiments around a 132M-parameter model, evaluating language modeling and downstream paraphrase detection performance.
- Investigated the fundamental trade-offs between model memory, computational efficiency, quantization, and quality for resource-constrained inference.
Voice AI Agent
Real-time conversational voice system for production phone calls
- Built a production-style real-time voice agent capable of handling phone conversations through speech recognition, LLM reasoning, and neural text-to-speech.
- Integrated Twilio, WebSockets, STT/TTS services, LiveKit-style real-time infrastructure, and FastAPI to create a low-latency conversational pipeline.
- Implemented streaming audio, call handoffs, interruption handling, session management, and debugging across the entire telephony/audio stack.
- Built versions using multiple voice providers, demonstrating the ability to reason about latency, audio formats, streaming, and vendor abstraction.
GTM Automation Platform
AI-powered sales infrastructure for prospecting and outreach at scale
- Built an end-to-end AI-powered GTM automation platform for identifying, enriching, scoring, and contacting business prospects.
- Engineered pipelines combining company discovery, website scraping, LinkedIn intelligence, Apollo organization search, Hunter domain discovery, and automated enrichment enabling AI-based voice calling outreach to different parts of the world for lead generation.
- Designed the system around FastAPI, PostgreSQL/Supabase, Docker, asynchronous pipelines, state management, and modular enrichment providers.
DocDynamo
2,600+ UsersVerification-first RAG platform — flagship product of OpenRAG Innovations (2024–2026)
- A startup product built around reliable, source-grounded generative AI, which grew out of an earlier document intelligence tool.
- Built a full RAG stack spanning document ingestion, retrieval, corrective context generation, web grounding, verification, trust scoring, and reliability analytics.
- The platform evolved from a PDF/YouTube research assistant into a broader de-hallucination and AI trust layer for AI applications.
- Scaled the product to thousands of users while building production APIs, vector-search pipelines, multimodal generation, and an Android/web experience.
AI Research Assistant — ResearchGuru
A precursor to OpenRAG's document intelligence and RAG ecosystem
- Built an AI research assistant designed to help researchers navigate technical literature, documents, and external information.
- Combined document understanding, retrieval, LLM reasoning, summarization, concept extraction, and web information discovery into a unified workflow.
- The system was designed around the practical research workflow rather than a generic chatbot, connecting papers → concepts → explanations → additional sources.
- This work became an important precursor to the broader document intelligence and RAG product ecosystem at OpenRAG.
Research Papers
TernaryLM: Memory-Efficient Language Modeling via Native 1.5-Bit Quantization with Adaptive Layer-wise Scaling
Status: Accepted · Camera Ready
GlobalSouthAI Workshop at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026)
Demonstrates native 1.5-bit quantization with per-layer precision scaling — enabling competitive LLM performance at extreme memory compression for edge deployment.
Optimization of Latent-Space Compression using Game-Theoretic Techniques for Transformer-Based Vector Search
Status: Accepted
A Transformer-Based Model for Enhanced Tokenization and Generation in Hindi Natural Language Processing
Status: Presented (Pending Publication)
Neural Orchestration for Multi-Agent Systems: A Deep Learning Framework for Optimal Agent Selection
Status: Presented (Pending Publication)
Model Context Protocol (MCP): A Lightweight, Modular Framework for Tool-Augmented LLM Agents
Status: Accepted (Camera Ready)
From Chalkboards to Chatbots: Exploring Teacher and Student Readiness for Agentic AI in Indian Schools
Status: Presented (Pending Publication)
Conversational Text Extraction with Large Language Models Using Retrieval-Augmented Systems
Status: Published
Innovative Fusion of LSTM and Bi-GRU Networks for Enhanced Hate Speech Detection in Social Media
Status: Published
Deep Learning in Industry 4.0: Transforming Manufacturing through Data-Driven Innovation
Status: Published
Advancements in Computer Vision and Machine Learning for Food Quality Evaluation: A Comprehensive Review for the Food Industry
Status: Published
Articles
2024
Beyond Reality: Virtual Worlds Shaping the Future of Food Science
Food Infotech Magazine — January 2024
Read Article →Advancements in Food Safety and Quality Evaluation Using Computer Vision and Machine Learning
Food, Marketing and Technology Magazine — March 2024
Read Article →Implementing the Technological Intelligence in Agricultural Produce
Food, Marketing and Technology Magazine — June 2024
Read Article →Edible and Biodegradable Packaging Innovations
Food, Marketing and Technology Magazine — December 2024
Read Article →2023
Two Layer Classifiers of MVS for Effective Grading, Safety & Quality Evaluation in Food Industry
Food Infotech Magazine — April 2023
Read Article →Crunching the number of safer foods: how big data is transforming food safety
Food Infotech Magazine — May 2023
Read Article →Cracking the Code! Unveiling the Hidden World of Food Safety with MicroWaves & ML
Food Infotech Magazine — June 2023
Read Article →Revolutionizing Agri-Food Supply Chain: Harnessing the Power of IoT
Food Infotech Magazine — July 2023
Read Article →Revolutionizing Food Safety: How BlockChain and Lighting Network are changing the game
Food Infotech Magazine — September 2023
Read Article →Revolutionizing Brain Tumor Detection: The CNN-based Medical Imaging Breakthrough
Informs London Magazine — December 2023
Read Article →Blog — Writing on Medium
Essays on AI systems, model releases, and industry analysis, published on Medium and Towards AI.
August 15, 2026
Understanding Latency in Multi Agentic AI Systems | DeepDive #1
Model latency is only 30–40% of the total latency shown by the system.
Read on Medium →
July 21, 2026
Kimi K3: 2.8T Most Capable Model by Moonshot.ai
Moonshot's flagship 2.8T-parameter model — novel attention and MoE design for efficiency.
Read on Medium →
June 9, 2026
The Complete Guide to Attention Variants in Transformers: From Scaled Dot-Product to Flash
A comprehensive overview of transformer attention mechanisms and their evolution.
Read on Towards AI →
May 5, 2026
The Hidden Engine Behind Fast LLM Inference | Why It's Necessary for All Wrapper Applications
Infrastructure optimizations powering rapid LLM responses beyond model speed alone.
Read on Medium →
April 19, 2026
The Quiet Revolution Happening Inside Your Software — and Why Most Developers Are Already Behind
How agentic AI systems are quietly reshaping software development.
Read on Medium →
April 11, 2026
How Perplexity AI Makes Money on Top of GPT, Claude, Gemini: Inside the Answer Engine Business
Analyzing a $21B startup's business model built on third-party LLMs.
Read on Medium →
April 2, 2026
Meta's TRIBE v2: The AI That Predicts How Your Brain Reacts to the World
A neuroscience foundation model trained on 500+ hours of fMRI data.
Read on Medium →
February 8, 2026
Kimi K2.5: The Next Milestone in Multimodal and Agentic AI
Moonshot's multimodal model advancing visual understanding and agentic capability.
Read on Medium →
January 25, 2026
Why LLMs Forget Long Conversations | AI Replacing Humans Narrative
Context window limits keep LLMs from developing genuine memory, unlike human cognition.
Read on Medium →
January 5, 2026
Compression Is Easy. Preserving Meaning Is Not | Compressed Language Models — Beginner 101
Shrinking the model is the easy part. Keeping it intelligent? That's hard.
Read on Medium →Authored Books
Fundamentals of Convolutional Neural Networks with TensorFlow
A deep dive into Computer Vision, Object Detection, AI, CNNs, and TensorFlow.
Available on: Amazon, Kindle, Flipkart
Get The Book
Trust in the Age of Agentic AI Economy
A C-suite guide to building trustworthy, reliable agentic AI systems at scale. Drawn from the production infrastructure built at OpenRAG Innovations, covering hallucination mitigation, retrieval reliability, multi-agent coordination, and the emerging trust frameworks enterprises need as AI systems become autonomous decision-makers.
Co-authored under OpenRAG Innovations Pvt. Ltd.
