I conduct research in Multimodal Deep Learning, Vision Transformers, and Computer Vision, and engineer high-performance, reproducible ML systems:

• Multimodal Transformers & Vision: Designing joint feature representation networks (BERT-BiGRU + ViT) for complex multimodal tasks like automated fake news detection.
• Production MLOps: Building modular pipelines tracking end-to-end data and model lineage using DVC, MLflow, Docker, and automated CI/CD deployment on AWS EC2.
• Hardware & Systems Analytics: Modeling LLM memory footprints, activation overhead, and distributed parallelism (TP / PP / DP).

Deep Learning • Multimodal Transformers • Springer IC2NS2 '26

• Architecture: End-to-end classification pipeline integrating BERT–BiGRU for contextual narrative semantics with a Vision Transformer (ViT-B/16) for patch-level visual representations.
• Empirical Outcome: Achieved 90.18% Test Accuracy, 0.9485 ROC-AUC, and 0.9084 Macro F1 across 30,900 samples on the Fakeddit benchmark.
• Stack: PyTorch BERT-Base BiGRU ViT-B/16 Feature Concatenation Fakeddit

MLOps • Containerization • CI/CD • Computer Vision

• Architecture: Production-ready automated image classification pipeline engineered for reproducible poultry disease diagnosis with modular data ingestion and validation pipelines.
• Infrastructure & Outcome: Attained 96.55% CV accuracy with version-controlled datasets via DVC, experiment tracking via MLflow, and automated deployment to AWS EC2 with Docker & FastAPI.
• Stack: DVC MLflow Docker FastAPI AWS EC2 GitHub Actions

Deep Learning Infrastructure • Memory Analytics • Serverless • [Live Tool]

• Architecture: Analytical hardware sizing engine designed to estimate exact GPU VRAM requirements and bandwidth bottlenecks for LLM training and inference.
• Precision Sizing: Models parameter memory footprint, dynamic KV cache, activation overhead, optimizer states (AdamW / FP16 / INT8 / FP8), and multi-GPU tensor/pipeline parallelism configurations (TP / PP / DP).
• Stack: Cloudflare Workers VRAM Analytics LLM Inference & Training Parallelism (TP/PP/DP) JavaScript ES6+

Web Platform • Chronological Analytics • High Performance • [Live Platform]

• Architecture: Algorithmic chronological analytics platform engineered for zero-dependency client computations, dynamic chart visualizations, and responsive glassmorphism UI.
• Performance: Microsecond client-side execution speeds with lightweight zero-bloat browser rendering and comprehensive semantic SEO optimization.
• Stack: JavaScript ES6+ Chronological Algorithms Glassmorphism UI SEO Architecture

@inproceedings{mishra2026unified,
  title={A Unified Multimodal Framework for Fake News Detection Using BERT and Vision Transformers},
  author={Mishra, Ayush and Kumar, Naveen},
  booktitle={Proceedings of the International Conference on Intelligent Computing, Cognitive Networks, and Smart Systems (IC2NS2 2026)},
  year={2026},
  note={Paper ID: 182; Accepted and Presented},
  publisher={Springer}
}