Gruhesh Sri Sai Karthik Kurra

• Morphology-Aware Nested Character Embeddings for Word Representation

  AcceptedIEEE ICCCMLA 2025 · 2025

  Character-level word representation: nested character sequences fused with hierarchical Transformers, CNNs, and BiLSTMs. +58.3 points on morphological-variant sentiment (91.67% vs 33.33% for Word2Vec), 99.0% on DBpedia and 93.5% on AG News classification, and 81.3% F1 on German NER (GermEval 2014). Morphology breakdown: 87.3% plurals, 85.9% past-tense, 81.5% superlatives, against 75.6% for FastText.

• BIE: Bit-Index Encoding for Efficient Neural Network Weight Compression

  AcceptedZenodo · September 2025

  Bit-level indexing for neural network weights: compression up to 40× with MSE < 10⁻⁶ and minimal accuracy drop. Open-source implementation with optimised sparse kernels.

• Hybrid RAG-Enhanced Deepfake Detection: Combining Retrieval-Augmented Generation with Visual Inconsistency Analysis

  AcceptedZenodo · August 2024

  Privacy-first deepfake detection: RAG semantic retrieval plus visual anomaly detection and an uncertainty-aware classifier.

• Voice-Activated AI for Seamless Computer Interaction

  AcceptedIJNRD · January 2025

  Voice command system with intent classification and system control for human–computer interaction.

• Global Remote RAM Sharing: A Novel Framework for Distributed Computational Systems

  AcceptedIJNRD · December 2024

  Remote RAM sharing for distributed systems under compute and memory constraints.

• Dynamic Auto-Finetuning of Language Models Based on Confidence-Driven Knowledge Integration

  AcceptedIJNRD · November 2024

  Confidence scores trigger selective fine-tuning without a full retrain, using confidence metrics and knowledge integration.

Under Review & Working Papers

• Prism Tuning: Repulsion-Trained Seed Embeddings in a Frozen Transformer for Non-Redundant Generation

  Under ReviewIEEE NEPCON 2026 · Submitted 2026

  K learnable seed vectors trained by output-level repulsion inside one frozen transformer, giving K non-redundant generation streams from a single prompt. Diversity is paid once at training, with zero overhead at inference, unlike generate-then-deduplicate. Uses a generated-position hidden-state surrogate for non-differentiable decoded text, corrects representation anisotropy in the repulsion loss, and splits each seed into shared-style / repelled-content parts to block style-only divergence.

• Dual-Stream Artifact Detection with Iterative Evidence Refinement for Frame-Level Deepfake Recognition

  Under ReviewIEEE NEPCON 2026 · Submitted 2026

  Boundary (Sobel-edge) and frequency (multi-band FFT) streams trained under an orthogonality penalty and fused via iterative cross-attention (ERM). Trained on 227,504 frames from FaceForensics++, Celeb-DF v2, and WildDeepfake. 96.3% AUC on the video-disjoint test split (87.1% on Celeb-DF alone), after an internal audit caught train/test video leakage in WildDeepfake and the paper cites the corrected number over the inflated 95.6% headline. Code, frame corpus, audit script, and trained checkpoint public: GitHub · Kaggle dataset.

• RADAR: Reasoning-Augmented Deepfake Artifact Recognition via Multi-Branch Evidence Aggregation

  Preprint2026

  Deepfake detection as structured evidence aggregation: three specialised branches (skin texture, boundary aliasing, AI-generation fingerprints) feed an iterative cross-attention reasoning module (R-Former) for per-artifact confidence scores and saliency maps, trained from random initialisation to remove ImageNet bias. Architecture and evaluation protocol; empirical validation is in the companion Dual-Stream study above.