At VelocitaiDigitalPrivateLimited, I improve computer-vision and AI proctoring systems. I refined OpenCV frame analysis to detect lighting conditions and reduce false lighting and absence alerts in live testing.
I combined multimodal vision-LLM outputs with MTCNN face detection to recover background faces missed by its size filter, with no false positives in testing. I also rebuilt FaceNet identity references from interview photos in AWS S3 and added timestamped storage keys to prevent recording mix-ups.
I enhanced CV and job description parsing with LLMs, RAG, regex, and databases, and improved existing speech-to-text and text-to-speech pipelines. I contributed to backend modularization and deployment using FastAPI, Docker, Git, GitHub Actions, and Argo CD.
In independent research, I designed and implemented a controlled LoRA fine-tuning pipeline for CLIP ViT-B/32 to study cross-modal image-text retrieval. My experiments achieved up to 9.74 percentage points improvement in mean Recall@1 on Flickr30k and 7.29 percentage points on COCO over frozen CLIP.
