Aman Kumar

About Me

I engineer highly robust, hardware-efficient predictive pipelines, specializing in Natural Language Processing (NLP), Computer Vision, and advanced sequence modeling. My primary objective is bridging the gap between theoretical deep learning and scalable, real-world deployment for high-stakes, computationally constrained environments.

Currently, I serve as Co-Founder and Lead AI Researcher at the Halo Mind Research Group, alongside dual AI Research Intern affiliations at the Indian Statistical Institute (ISI) Kolkata and the University of Salford (UK). I direct the formulation and code implementation of complex architectures targeting:

Space AI & Remote Sensing: Engineering the DDV-GNet architecture for ultra-high-throughput aerospace defect detection (IEEE SPACE 2026) and Deep Delta Vision Mamba for lightweight satellite land-cover classification (IEEE CONECCT 2026), alongside real-time CNN-DDL frameworks for ionospheric TEC forecasting.

Clinical NLP & Medical Vision: Leading neural extraction pipelines for complex clinical case report summarization (CLEF 2026 MultiClinSum-2) and developing robust brain tumor classification models using Attention-Enhanced Swin Transformers and SafeMed-SSL for reliable diagnostic pipelines.

Cyber-Physical Resilience & Multimodal OCR: Formulating Intrinsic Neural Firewalls for adversarial anomaly rejection (Oral Presentation, WIN 6.0 2026) and engineering self-supervised, end-to-end CV/NLP architectures for digitizing degraded ancient Ashokan Brahmi scripts.

With a portfolio of multiple accepted manuscripts across IEEE, Springer, and Taylor & Francis venues, my workflow is strictly driven by rigorous ablation, architectural optimization, mathematically sound uncertainty quantification (Conformal Prediction), and open-source reproducibility. I am currently advancing foundational AI representations in preparation for international Master’s or Direct Ph.D. research for the Mid-2027 intake.

Experience

Research Intern (Computer Vision & Deep Learning)
Indian Statistical Institute (ISI), Kolkata | Hybrid | Jun 2026 – Present

• Conducting research in complex spatial analysis and image restoration under the supervision of Prof. Umapada Pal, bypassing traditional segmentation bottlenecks.

Research Intern (Deep Learning & Sequence Modeling)
The University of Salford, UK | Remote | Jun 2026 – Present

• Collaborating internationally on deep learning architectures for text recognition and visual sequence modeling frameworks under the supervision of Dr. Shivakumara Palaiahnakote

Undergraduate AI Researcher & Team Lead
Sikkim Manipal Institute of Technology (SMIT) | On-site | Aug 2024 – Present

• Directing multi-disciplinary deep learning research spanning medical image segmentation, remote sensing, and NLP-driven document analysis.
• Integrating advanced statistical frameworks (Conformal Prediction, Deep Delta Learning) to establish provable reliability and uncertainty quantification in complex physical and visual systems.

Core Research Themes

• Medical Image Analysis & Diagnostics: Optimizing deep learning pipelines for clinical reliability, including brain tumor classification, domain-equalized spatial-frequency fusion for synthetic dermatology detection, and safe semi-supervised learning for malaria diagnosis.
• Edge AI & Hardware Optimization: Engineering low-latency, high-throughput convolutional networks (like DDV-GNet) optimized specifically for constrained IoT devices and space manufacturing hardware.
• Cyber-Physical Systems (CPS) Security: Developing robust anomaly rejection frameworks and intrinsic neural firewalls to protect critical industrial infrastructure from False Data Injection Attacks without relying on post-hoc detectors.
• Uncertainty Quantification & Remote Sensing: Integrating Marginal Split Conformal Prediction into deep learning architectures to establish rigorous, distribution-free reliability in high-stakes domains like high-resolution urban change detection and ionospheric TEC forecasting.
• NLP & Computational Linguistics: Utilizing BERTopic, SBERT, and advanced regression statistics (OLS, Chi-Square) to quantify long-term semantic evolution, socio-emotional shifts, and human behavior within an 89K+ document corpus.
• Computer Vision & OCR: Designing complete ETL pipelines and end-to-end architectures leveraging SimCLR and WGAN-GP for degraded ancient Brahmi script recognition.

Independent Research Group: Halo Mind

Hybrid Architectures & Lightweight Optimization | Machine Intelligence & Neural Dynamics

I am the Lead ML Systems Engineer and co-founder of Halo Mind, an independent, interdisciplinary research group alongside Lead Research Architect Latchan Chhetri. We bridge continuous physical dynamics with discrete neural computation, engineering hardware-efficient neural mechanisms like sub-quadratic State Space Models (Mamba) and Deep Delta Learning (DDL) to process high-throughput, volatile data streams.

Operating outside traditional, heavily funded academic pipelines, our methodology is strictly driven by absolute resource constraints. This environment forces rigorous architectural innovation, exhaustive ablation, and mathematically bounded efficiency.

My Role: Lead ML Systems Engineer

• Focus: Pipeline Optimization, Conformal Prediction, Applied Frameworks.
• Execution: I translate theoretical architectures into robust, hardware-optimized code. I manage complex spatial data pipelines and integrate advanced statistical frameworks to establish provable reliability and distribution-free uncertainty quantification.

Our applied domains span extreme space weather forecasting (Ionospheric TEC), robust medical image analysis (Brain Tumor & Synthetic Dermatology), and low-resource sequence modeling for ancient epigraphy (Ashokan Brahmi). Prioritizing mathematical truth over brute-force computation, we design self-contained, uncertainty-aware frameworks capable of autonomous anomaly rejection and real-time distribution shift adaptation.

Visit our Research Group at: halomind-research.github.io
Contact: halomind.research.group@gmail.com

---