publications

This paper presents a real-time assistive system for outdoor navigation of visually impaired users in complex Indian urban environments. A BLIP vision-language model is fine-tuned through a 3-stage LoRA training strategy on a custom 427-scene dataset built from Google Street View API imagery, yielding a +15.6% improvement across BLEU, METEOR, ROUGE and semantic similarity metrics. The system is deployed on a Raspberry Pi through an edge-cloud hybrid architecture, and covers the full research pipeline: dataset curation, model training, evaluation framework and edge deployment. The trained model is released on HuggingFace and the code is open-sourced.

This work proposes paraphrase variance across K=5 semantic-preserving rewrites as a label-free, unsupervised proxy for hallucination in large language model outputs, requiring no ground-truth answers at inference time. The signal is evaluated using BERTScore, NLI contradiction rate and AUC-ROC on the TriviaQA and Natural Questions benchmarks. Experiments run on Llama-3-8B-Instruct (BF16, FlashAttention-2) on 2x NVIDIA L40S, and extend to Llama-3-70B for a cross-scale validation claim.