Hi there, I'm Nandakishor Mukkunnoth 👋

Founder & CEO at ConvAI Innovations | AI Researcher & Systems Engineer

Building explainable, privacy-preserving, and ultra-low-latency AI models for healthcare, edge systems, and calibrated decision-making.

I am an AI researcher, engineer, and entrepreneur passionate about bridging the gap between frontier AI research and real-world societal impact. My work centers on:

• Offline, Privacy-First Healthcare AI: Developing medical intelligence systems that run 100% locally on mid-level PCs to serve rural clinics without cloud dependency or patient data egress.
• Calibrated "System 1" Decision Engines: Building fast, non-autoregressive decision models that replace bulky LLMs for structured classification, routing, and guardrails in under 40 milliseconds.
• Explainable Machine Learning: Creating attention heatmaps and semantic state-projection techniques to dismantle the "black box" nature of neural networks in clinical diagnosis.

• The Mission: In rural primary health centers, cardiologists are often unavailable during critical golden-hour windows. Acute cardiac events (STEMI) cause ~228 deaths every hour globally.
• The Solution: A completely offline desktop application that processes 12-lead ECG images and blood reports. Powered by a fine-tuned MedGemma 4B multimodal model (trained on 1M+ ECGs via 8×A100s on India's AIRAWAT supercomputer during the C-DAC AI Innovation Challenge) coupled with an ONNX-quantized MedSigLIP vision encoder for instant attention heatmap overlays.
• Stack: Electron, node-llama-cpp (GGUF), ONNX Runtime, LangGraph JS, SQLite. Zero cloud egress for strict HIPAA/GDPR compliance.
• Desktop App Repository

• The Mission: Stop using slow, hallucinatory 8B+ LLMs for simple reflex decisions (routing, prompt guardrails, triage).
• The Solution: A 421M-parameter bidirectional decision model (ModernBERT-large + scratch decision head) trained via RLCD (Reinforcement Learning for Calibrated Decisions) using strictly proper scoring rules (log-score + spherical score + ranked probability score).
• Key Features: Evaluates choice, score, and noul (calibrated probabilities) in a single forward pass (~33–38 ms on GPU). Delivers 92.2% accuracy at 50% coverage for selective automation.
• PyPI Package · Hugging Face Model · Interactive Space · Dev.to Article

• A lightweight Python framework designed to reduce LLM hallucinations by projecting internal hidden states into a semantic projection space and measuring alignment with the input prompt before token generation.

• Real-time, completely private sales call analysis running locally via Qwen3-ASR and Qwen3.5 GGUF through llama.cpp in an Electron shell. Zero audio or transcript data leaves the machine.

• Custom Hindi embedding models built from scratch for multilingual Retrieval-Augmented Generation (convaiinnovations/hindi-embeddings).

1. Confidence-Aware Routing for Large Language Model Reliability Enhancement: A Multi-Signal Approach to Pre-Generation Hallucination Mitigation
   Nandakishor M · arXiv:2510.01237 (2025)

2. SalesRLAgent: A Reinforcement Learning Approach for Real-Time Sales Conversion Prediction and Optimization
   Nandakishor M · arXiv:2503.23303 (2025)

3. ForcePose: A Deep Learning Approach for Force Calculation Based on Action Recognition Using MediaPipe Pose Estimation Combined with Object Detection
   Nandakishor M et al. · arXiv:2503.22363 (2025)

4. DeepRAG: Building a Custom Hindi Embedding Model for Retrieval Augmented Generation from Scratch
   Nandakishor M · arXiv:2503.08213 (2025)

5. Continuous Learning Conversational AI: A Personalized Agent Framework via A2C Reinforcement Learning
   Nandakishor M, Dr. Anjali M · arXiv:2502.12876 (2025)

6. High-Accuracy ECG Image Interpretation using Parameter-Efficient LoRA Fine-Tuning with Multimodal LLaMA 3.2
   Nandakishor M, Dr. Anjali M · arXiv:2501.18670 (2025)

• Democratizing Clinical Diagnostics: Partnering with medical professionals (like Dr. Anjali M, Assistant Professor at Dr. Moopan's Medical College) to curate authentic clinical archives and validate AI systems for community healthcare workers.
• Patient Privacy & Edge Sovereignty: Championing offline AI models that run without internet connectivity, ensuring full compliance with HIPAA and GDPR while functioning reliably in bandwidth-constrained rural hospitals.
• Genuine Open Science: Publishing code, model weights, and datasets openly to empower developers and independent researchers worldwide.

• Languages: Python, JavaScript (ES6+), TypeScript, C++, SQL
• AI / ML Frameworks: PyTorch, Hugging Face Transformers, Reinforcement Learning (PPO, A2C, RLCD), ONNX Runtime, llama.cpp (GGUF), MediaPipe
• Application & Edge Engineering: Electron, LangGraph, Node.js, FastAPI, Gradio
• High-Performance Computing: SLURM Cluster Orchestration, C-DAC AIRAWAT Supercomputing, Distributed Data Parallel (DDP)

If my open-source models, healthcare diagnostic tools, or research papers have helped you, consider supporting my independent research: