Researcher in Natural Language Processing and AI

About Me

Hi, I’m Mihir Panchal, an aspiring researcher exploring Natural Language Processing, Large Language Models and Graph Representation Learning. I have collaborated with leading institutions including Georgia Tech, King’s College London, National University of Singapore, IIT (ISM) Dhanbad, IIT Jodhpur, IIT Mandi, IIT Patna, DRDO, and NIDM, often bridging theoretical methods and real world applications. I enjoy collaborating with researchers across disciplines and geographies and am always open to new opportunities that push the boundaries of AI and human machine collaboration. Below is a non exhaustive list of my current research interests. Feel free to reach out via your favorite platform :)

Research Interests:

Mechanistic Interpretability of Reasoning Techniques in LLMs
Memory Augmented LLM Architectures and Retrieval Augmented Generation (RAG) for Contextual Reasoning
Multi-Agent Collaboration Frameworks for Evaluating and Enhancing LLM Reasoning
Simulation Based Interpretability for Sequential Language Understanding

Research Collaborations

Georgia Institute of Technology & NIDM India

Generative Modeling for Disaster Management | Georgia, Atlanta

August 2025 – Present

Developed CC-GRMAS, a multi agent graph neural system for spatiotemporal landslide risk forecasting, integrating GNNs, Retrieval-Augmented Generation, and graph databases for proactive disaster preparedness.
Collaborated with a cross institutional team to preprocess NASA Global Landslide Catalog, design spatial graphs, and evaluate graph based models achieving 0.79+ F1-score while reducing model complexity by 99.9%.
Collaborated with Dr. Ying-Jung Chen (Georgia Tech) and Dr. Surya Parkash (NIDM India) on cross institutional disaster management research.

King's College London & National University of Singapore

Reasoning and Interpretability of LLMs | Singapore

June 2025 – Present

Advanced multi-lingual LLM reasoning and interpretability, leveraging reasoning chains, circuit tracing, logit lens, and attention head analysis to enhance model transparency and performance across diverse linguistic tasks.
Conducted systematic probing over model internals by recording activations across residual streams, attention layers, attention output projections, and MLP blocks to trace information flow and identify causal pathways.
Collaborated with Dr. Mamta Sahni (King’s College London) and Dr. Deeksha Varshney (NUS) on multilingual AI interpretability research.

IIT (ISM) Dhanbad & IIT Jodhpur

Graph Representation Learning in LLMs | Jodhpur, India

June 2025 – Present

Optimized Graph of Thoughts reasoning using Block Successive Upper bound Minimization and graph coarsening on GSM8K dataset, finetuning models to improve scalability and efficiency for complex reasoning tasks in LLMs.
Proposed a graph structured reasoning framework that abstracts reasoning chains into compact representations, enabling models to perform multi step reasoning without explicit prompting or additional finetuning.
Collaborated with Dr. Manoj Kumar (IIT ISM Dhanbad) and Dr. Deeksha Varshney (IIT Jodhpur) on graph based reasoning in large language models.

DRDO & IIT Mandi

Speech Processing and NLP | Mandi, India

January 2025 – September 2025

Implemented ECAPA-TDNN architecture incorporating SE-Res2Blocks, channel attention, and skip connections for extracting high quality speaker embeddings, deployed by Government of India in multilingual systems.
Developed end-to-end pipelines for speech preprocessing, embedding extraction, and diarization scoring, ensuring accurate segmentation of overlapping speakers in diverse Indian multilingual datasets.
Collaborated with Dr. Prabhat Bharti (IIT Mandi) and Dr. Asif Ekbal on government deployed speech processing systems for multilingual Indian contexts.

AI-NLP-ML Research Group (IIT Patna)

Computational Linguistics in Peer Review | Bihar, India

December 2023 – September 2025

Implemented GraphRAG to assess peer review consistency, leveraging knowledge graphs and counterfactual reasoning at word, sentence, and aspect levels to improve transparency and reliability in review processes.
Developed multi task models with 6 ablation variants, comprising SciBERT, BERT-Base, and Bi-LSTM, connected to multiple attention blocks to determine aspect categories and their sentiments for any given review.
Collaborated with Mr. Viral Dalal (NYU) and Dr. Mayank Agarwal (IIT Patna) on computational linguistics approaches to scientific peer review analysis.

Research Publications

CC-GRMAS - A Multi-Agent Graph Neural System for Spatiotemporal Landslide Risk Assessment in High Mountain Asia

Neural Information Processing Systems (NeurIPS) 2025 - In Review

Mihir Panchal, Ying-Jung Chen, Surya Parkash

Co-authored a novel framework integrating GNNs and multi agent coordination for climate resilient disaster preparedness, validated on NASA’s Global Landslide Catalog in High Mountain Asia (HMA region).

Indic-TunedLens - Interpreting Multilingual Models in Indian Languages

International Joint Conference on Natural Language Processing (IJCNLP AACL) - Minor Revision

Mihir Panchal, Deeksha Varshney, Mamta Sahni, Asif Ekbal

Developed Indic-TunedLens framework for multilingual LLM interpretability in Hindi, Marathi & Bengali, using language specific affine transformations to improve layer wise decoding accuracy over English centric methods.

Not all peers are significant - A Dataset Exhaustive vs Trivial Scientific Peer Reviews Leveraging Chain-of-Thought Reasoning

Scientometrics, Springer - Accepted

Prabhat Kumar Bharti, Mihir Panchal, Viral Dalal, Mayank Agarwal, Asif Ekbal

Constructed InsightfulPeer, a peer review classification dataset leveraging Chain-of-Thought reasoning with LLMs (LlaMa-3.1, GPT-4, Mixtral, Gemma) and evaluated model human agreement using Cohen’s and Fleiss Kappa.

ConsistentPeer - Reviewers Through GraphRAG-Driven Counterfactuals to Measure Consistency in Peer Review

International Journal of Data Science and Analytics, Springer - Minor Revision

Prabhat Kumar Bharti, Mihir Panchal, Viral Dalal

Proposed a novel knowledge graph based framework to evaluate consistency in peer reviews by correlating review text, confidence scores, and ratings, incorporating counterfactual reasoning to detect and resolve inconsistencies.

LEDGE - Leveraging Dependency Graphs for Enhanced Context Aware Documentation Generation

Automated Software Engineering, Springer - Preprint

Mihir Panchal, Arnav Deo, Prinkal Doshi, Varad Prabhu, Chetashri Bhadane

https://www.researchsquare.com/article/rs-6827966/v1

Researched and developed LEDGE, a novel framework integrating dependency graphs and GraphRAG to advance automated, context aware software documentation generation, empirically validated through rigorous evaluation.

Co-Reviewer - Are LLMs on the Same Page as Human Reviewers? An Agentic AI Framework for Evaluating Review Quality and Consensus

Scientometrics, Springer - Minor Revision

Prabhat Kumar Bharti, Viral Dalal, Mihir Panchal, Mayank Agarwal, Asif Ekbal

Designed Co-Reviewer, a multi agent LLM framework for collaborative peer review generation and refinement; evaluated its outputs against human reviews across informativeness, sentiment, and editorial decision alignment.

PeerGauge - a Dataset for Peer Review Disagreement and Severity Gauge

Language Resources and Evaluation, Springer - In Review

Prabhat Kumar Bharti, Mihir Panchal, Viral Dalal

Built PeerGauge, a 28k-pair peer review contradiction severity dataset (-5 to +5 scale) from ICLR/NeurIPS, and benchmarked models (RoBERTa, LSTM, SciBERT, LLMs) for automated severity prediction.

Game Machine and Algorithm towards Trends in Game States using Machine Learning and Deep Learning

2023 10th International Conference on Computing for Sustainable Global Development (INDIACom) - Published

Mihir Panchal, Chintan Jagdish Dodia, Pankaj Dulabhai Rathod

https://ieeexplore.ieee.org/abstract/document/10112350

Developed a hybrid approach integrating machine learning and deep learning, utilizing Markov Chain models to analyze game states in chess, poker, cricket, and football.

Teaching Experience

Why Research Papers? A Comprehensive Guide | DJSACM Research Seminar This comprehensive seminar was delivered to 200+ computer science students, covering research methodology, literature review, research question formulation, and academic writing best practices. The workshop featured interactive modules on critical analysis of research papers, identifying research gaps, and hands-on guidance with peer review, citation management, and publication venue selection. Watch the recording here
Transformers in AI - From Perceptrons to Generative Models | SBMP Alma Mater This workshop was conducted for 40+ polytechnic students on the evolution of AI architectures, covering perceptrons, neural networks, attention mechanisms, and modern transformer models (BERT, GPT). The curriculum included hands-on coding sessions in Python, TensorFlow, and Hugging Face, enabling students to build, fine-tune, and apply transformer models for NLP tasks. View the LinkedIn post here
Building Intelligent Applications with LangChain and LLMs | TFUG ML Mumbai This workshop was led for 300+ ML practitioners on building production-ready AI applications with the LangChain framework, integrating multiple LLMs with custom data sources and business logic. Advanced concepts such as retrieval-augmented generation (RAG), vector databases, prompt engineering, and chain composition were covered with hands-on examples of chatbots, QA systems, and content pipelines. View the LinkedIn post here
Leveraging Boto3 - Pythonic access to S3 and SNS on AWS | Mumpy June Meetup 2024 This hands-on workshop was delivered to 50+ Python developers on AWS integration with Boto3 SDK, covering S3 operations, SNS messaging, IAM security, and cloud best practices. Participants were guided through coding exercises on building scalable data pipelines, automated backups, and real-time notification systems with AWS. View the LinkedIn post here
From Code to Community - Publishing PyPI Packages | FOSS United June Meetup 2024 This workshop was presented to 50+ developers on Python package development and distribution, covering PyPI publishing workflow, versioning, documentation, and CI/CD integration. Open source best practices were demonstrated through live coding of a sample package, showcasing issue tracking, contribution guidelines, and community engagement strategies. View the LinkedIn post here
Unlocking the Power of Computer Vision with Mediapipe | Mumpy March Meetup 2024 This interactive workshop was conducted for 40+ developers on computer vision using Google’s MediaPipe, covering face detection, pose estimation, hand tracking, and gesture recognition. Hands-on coding sessions were led to build real-time applications such as fitness trackers, and interactive games with cross-platform deployment. View the LinkedIn post here
Mastering Python - A Hands-on Workshop for Developers | NMIMS Navi Mumbai This intensive 3-hour workshop was delivered to 80+ engineering students, covering Python fundamentals to advanced topics such as OOP, file handling, and key libraries (NumPy, Pandas, Matplotlib). Students were guided through hands-on projects in data analysis, web scraping, API development, and automation, supported with Jupyter notebooks, exercises, and project templates. View the LinkedIn post here

Education

University of Mumbai - Dwarkadas Jivanlal Sanghvi College of Engineering

Bachelor of Technology in Computer Engineering, Honors in Intelligent Computing | CGPA 9.00/10.0

August 2023 – May 2026

The curriculum encompasses foundational courses in Data Structures & Algorithms, Operating Systems, Database Management Systems, and Computer Networks, which emphasize algorithmic complexity analysis, concurrency control, distributed query processing, and network protocols.
Advanced training in Artificial Intelligence, Machine Learning, Deep Learning, and Natural Language Processing is complemented by specialized modules in Computer Vision, Speech Processing, and Graph Neural Networks, with a strong emphasis on optimization techniques, large scale model training, and research methodology.
Complementary coursework in Engineering Mathematics, Probability & Statistics, and Linear Algebra, along with systems oriented subjects like Computer Architecture, Digital Logic Design, and Microcontrollers, provides the mathematical rigor and hardware level insights essential for scalable and intelligent computing systems.

Shri Bhagubhai Mafatlal Polytechnic and College of Engineering

Diploma in Information Technology | CGPA 9.45/10.0

August 2020 – May 2023

The curriculum emphasizes Programming and Software Development through C, C++, Java, Python, SQL, and Web/Mobile technologies, with practical training in database driven applications, modular design, and system deployment.
Core subjects in Computer Networks, Operating Systems, Data Structures, and Software Engineering are complemented by System Analysis and Design, focusing on distributed architectures, process synchronization, data organization, and lifecycle modeling of large scale systems.
To bridge hardware and applied intelligence, I gained experience in Digital Electronics, Microprocessor Programming, and Computer Architecture, reinforced by Mathematics, Statistics, and Data Analytics. This culminated in a final year project on autonomous drone systems, integrating embedded programming, real time communication, and database driven control for aerial navigation and monitoring applications.

Professional Experience

Infiheal Healthcare Private Limited

Software Engineering Intern | Mumbai, India

June 2024 – August 2024

Boosted model accuracy by 15% using advanced AI, improved data processing efficiency by 20% with classification models, and built robust mental health AI apps using FastAPI, Docker, Git, and AWS for cloud deployment.
Utilized AWS services to improve system reliability by 25%, cut infrastructure costs by 30%, automated workflows achieving a 40% reduction in processing time, and collaborated with multidisciplinary teams in agile environments.
Collaborated with Ms. Srishti Srivastava and Mr. Vipul Chivate on mental health AI applications and cloud infrastructure optimization.

Skillsvista

Full Stack Developer Intern | Mallow, County Cork (Remote)

July 2022 – September 2022

Architected responsive web components using modern frontend technologies, resulting in a 30% improvement in page load times and a 20% increase in user engagement, leading to better user experience and retention.
Engineered highly efficient RESTful APIs and data models in Django, achieving 99.9% uptime and reducing API response times by 40%, improving overall application performance and scalability.
Collaborated with Mr. Shrenik Doshi and Mr. Jason Culloty on full stack web development and API optimization projects.

Fellowships & Volunteering

SimPPL Fellowship | Mozilla Funded Program This fellowship involved applying knowledge graph driven representation learning with GraphRAG to generate enriched embeddings that integrate syntax, semantics, and structural properties of large scale software systems. The work utilized static code analysis to create parsers for software, generating embeddings that capture dependency structures, control flows, and semantic patterns for downstream graph based analysis. The project focused on advancing AI driven software analysis through innovative graph neural network approaches. View project details here
Research Head | Association for Computing Machinery DJSCE In this leadership role, mentored 50+ students in research methodologies and technical writing, fostering academic excellence and research culture within the computer science community. Conducted 5+ comprehensive workshops on research paper writing, publishing strategies, and core technical skills to enhance academic and professional development. The initiative significantly improved student’s research capabilities and publication success rates across various technical domains.
Community Volunteer | FOSS United Mumbai Actively volunteered for FOSS United Mumbai community events, supporting the organization and execution of open source technology meetups and conferences. Contributed to event planning, speaker coordination, and community engagement activities that promote free and open source software adoption. The volunteer work helped build stronger connections within the Mumbai tech community and facilitated knowledge sharing among developers and technologists. Learn more about FOSS United Mumbai
Open Source Developer | Hacktoberfest 2022, 2023, 2024 Contributed to 15+ open source projects, strengthening global developer communities and promoting collaborative software development practices. Earned 3 Hacktoberfest completion badges for consistent, impactful contributions across diverse repositories and tech stacks. The contributions spanned multiple programming languages and frameworks, demonstrating versatility and commitment to the open source ecosystem. View contribution badges here

Blogs & Articles

What is SageMaker in AWS? This article provides an in depth exploration of Amazon SageMaker, AWS’s fully managed machine learning platform. It covers SageMaker’s core features including data preparation, model training, deployment, and monitoring capabilities. The guide explains how SageMaker simplifies the ML workflow by providing pre built algorithms, notebook instances, and automated model tuning features. Read the blog here
Everything You Need to Know About Open Source Development This comprehensive guide covers the fundamentals of open source development, including version control with Git, collaborative workflows, and best practices for contributing to open source projects. The article explores licensing models, community guidelines, and practical steps for developers to start their open source journey. Read the blog here
How to Upload JSON File to Amazon DynamoDB using Python? This tutorial demonstrates how to efficiently upload JSON data to Amazon DynamoDB using Python’s Boto3 SDK. It covers table creation, data preprocessing, batch operations, and error handling techniques. The guide includes practical examples for handling large datasets and optimizing upload performance for production environments. Read the blog here
Text to Speech by using ttsvoice - Python This article explores the implementation of text-to-speech functionality using Python’s ttsvoice library. It covers installation, basic usage, voice customization options, and integration with different applications. The guide provides practical examples for creating voice enabled applications and handling multiple languages and accents. Read the blog here
Creating ChatGPT Clone in Python This comprehensive tutorial walks through the process of creating a ChatGPT like conversational AI using Python. It covers natural language processing techniques, API integration, user interface design, and deployment strategies. The guide includes practical implementation details for building responsive chatbots with memory and context awareness. Read the blog here
Difference Between JPEG and MPEG This technical comparison explores the fundamental differences between JPEG and MPEG compression standards. It covers compression algorithms, use cases, quality considerations, and application scenarios. The article provides insights into when to use each format and their respective advantages in digital media processing. Read the blog here
Difference Between Digital Audio and MIDI This article examines the key differences between digital audio and MIDI (Musical Instrument Digital Interface) technologies. It explores data representation, file sizes, editing capabilities, and practical applications in music production and audio processing. The guide helps readers understand when to use each technology for different audio projects. Read the blog here
Hybrid Programming using Python and Dart This technical guide explores hybrid programming approaches using Python and Dart for cross platform application development. It covers API integration, data exchange mechanisms, and architecture patterns for building scalable applications. The article provides practical examples for combining Python’s data processing capabilities with Dart’s UI frameworks like Flutter. Read the blog here

Acknowledgements

Semantic Scholar API Gallery - CiteSpy CiteSpy was featured in Semantic Scholar’s official API Gallery for innovative academic data analysis and citation tracking. The tool has achieved 200+ downloads, demonstrating its value for bibliometric analysis and citation discovery. View the project feature here
GeeksforGeeks Technical Writing Contributor Contributed technical content to GeeksforGeeks with 55,000+ views across articles covering algorithms, data structures, programming languages, and emerging technologies, helping thousands of learners worldwide. View contributions profile here

Technical Skills

Languages: Python, C, C++, CUDA, SQL, LaTeX, Javascript, Typescript, Dart

Dev Tools: Linux, AWS, GCP, Docker, Kubernetes, Hugging Face, Wandb, Overleaf, GitHub

Frameworks: TensorFlow, PyTorch, FastAPI, Flask, Django, OpenCV, Langchain, React, React Native

More Projects & Articles

You can find my Research Resume at: research-portfolio
You can find more of my projects and Medium articles at: extended-portfolio