Gruhesh Sri Sai Karthik Kurra

• Morphology-Aware Nested Character Embeddings for Word Representation
  Accepted

  IEEE ICCCMLA 2025 · 2025

  Novel character-level representation modelling words as nested character sequences via bidirectional LSTMs and Transformers. Achieved +2–8% improvement on English classification tasks and +58.3% improvement on German NER over baselines.

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

  Zenodo · September 2025

  Novel compression framework representing neural network weights through bit-level indexing, achieving compression ratios up to 40× with MSE < 10⁻⁶ and minimal accuracy degradation. Includes complete open-source implementation with optimised sparse kernels.

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

  Zenodo · August 2024

  Privacy-first deepfake detection combining RAG-based semantic retrieval with visual anomaly detection and uncertainty-aware classification for enhanced accuracy and interpretability.

• Voice-Activated AI for Seamless Computer Interaction
  Accepted

  IJNRD · January 2025

  End-to-end voice command system with intent classification and system control for seamless human–computer interaction.

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

  IJNRD · December 2024

  Innovative framework optimising memory utilisation in distributed systems through global remote RAM sharing mechanisms for compute-constrained environments.

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

  IJNRD · November 2024

  Confidence score triggers selective fine-tuning without full retraining — novel approach for dynamically adapting language models using confidence metrics and knowledge integration strategies.

Preprints & Working Papers

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

  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 paid once at training, zero overhead at inference, unlike generate-then-deduplicate. Resolves the non-differentiability of decoded text with a generated-position hidden-state surrogate, corrects representation anisotropy inside the repulsion loss, and splits each seed into a shared-style / repelled-content part to block style-only divergence.

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

  2026

  Deepfake detection as structured evidence aggregation rather than end-to-end classification: three specialised branches — skin-texture, boundary-aliasing, and AI-generation-fingerprint analysis — feed an iterative cross-attention reasoning module (R-Former) for interpretable per-artifact confidence scores and saliency maps, trained entirely from random initialisation to remove ImageNet bias. Architecture and evaluation protocol; empirical validation carried out in the companion study below.

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

  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), reported after an internal audit caught train/test video leakage in the WildDeepfake portion and the paper cites the corrected number over the inflated 95.6% headline. Code, frame corpus, audit script, and trained checkpoint public.