At Aiverbalyze Technologies Private Limited, I trained and fine-tuned a multilingual Indic ASR system across 12 languages, achieving an average OIWER of 19.2%. I also curated synthetic multilingual datasets and used masked language-token conditioning to improve Malayalam ASR performance by approximately 40%.
I enhanced the existing speech-to-text inference pipeline with on-the-fly audio preprocessing, DeepFilterNet, Silero VAD v6, and utterance segmentation. I also improved the voice-to-voice pipeline with barge-in and interruption handling.
At BigThinx, I built a fashion computer vision pipeline for foreground segmentation and pose estimation from two-view images. I enhanced low-resolution and low-light inputs, improving existing pipeline performance by 30%, and deployed the system on a VM with a Gradio interface.
At IF MedTech, I fine-tuned and quantized a Vision Transformer for medical imaging, reaching 89% accuracy. I also built medical image workflows and implemented a medical RAG application; in my projects, I fine-tuned Qwen3-4B for text-to-SQL and developed a document-querying RAG system.

