Machine Learning · Secure AI · Brain Connectivity
Building machine learning systems that are not just accurate — but private, secure, and trustworthy. Research Intern at IIT BHU & Undergraduate Researcher at SRMIST.
Hi, I'm Aadi.
AI/ML Engineer · Security Researcher
I'm a 2nd-year B.Tech student in Computer Science (AI/ML) at SRMIST, working as a Research Intern at IIT BHU. My work sits at the intersection of machine learning, security, and privacy — building systems that are both intelligent and trustworthy.
I published my first paper as first author at IEEE ESCI 2026, applying Graph Attention Networks to autism detection on fMRI data. I also work on secure inference pipelines using Homomorphic Encryption at IIT BHU under Dr. Harsh Kashyap.
Research Intern — Trustworthy AI & Security
Indian Institute of Technology (BHU)Working on secure ML inference pipelines using Homomorphic Encryption (TenSEAL / CKKS). Focused on reducing the latency gap between encrypted and plaintext deep learning inference under Dr. Harsh Kashyap.
- ▹Secure inference pipelines with FHE (TenSEAL, CKKS)
- ▹Optimizing latency vs. security trade-offs in cloud ML
- ▹Designing Secure Aggregation protocols
Undergraduate Researcher
SRMIST — CINTEL LabFirst author of the DST-GAT framework for autism detection using fMRI brain connectivity data. Achieved state-of-the-art results on the ABIDE I benchmark.
- ▹First-authored paper accepted at IEEE ESCI 2026
- ▹AUC 0.74 on ABIDE I — outperforming all prior baselines
- ▹Dynamic Spatio-Temporal Graph Attention Network (DST-GAT)
DST-GAT: Dynamic Spatio-Temporal Graph Attention Network for Autism Spectrum Disorder Detection
We propose DST-GAT, a novel framework that models dynamic functional brain connectivity from fMRI time-series using Graph Attention Networks with temporal gating. Evaluated on the ABIDE I dataset, our approach achieves an AUC of 0.74, outperforming all prior baselines for ASD detection.
Network- and Power-aware Federated Learning on Raspberry Pi Edge Nodes: An Empirical Study
An empirical study of Federated Learning on constrained Raspberry Pi 4 edge nodes using containerized client emulation. We investigate the interplay between network conditions (simulated via Linux traffic control) and system resource utilization — CPU load, thermal throttling, and execution time — under FedAvg. Our results demonstrate how high-latency edge environments and thermal constraints non-linearly impact FL convergence, offering practical guidelines for deploying sustainable, network-aware Edge AI.
NeuraTwin
AI Mental Wellness Platform
Adaptive mental wellness platform featuring AURA — an emotion-aware conversational AI that adjusts its responses based on real-time sentiment analysis and user behavioral patterns. Built with Firebase and React.
NashGrid
Game Theoretic Security Simulator
Interactive visualization of Defender-Attacker strategies using Nash Equilibrium and Minimax search. Models adversarial scenarios for security analysis in real time.
Oneforall
WealthIQ FinTech — Hackathon
Hackathon submission delivering AI-powered personal finance intelligence with real-time portfolio insights.
ContextStream
Sub-500ms RAG Pipeline
Low-latency Retrieval-Augmented Generation system. Combines Next.js frontend, FastAPI backend, real-time WebSockets, and Pinecone vector search for semantic context delivery.
STACK