~ Developing a configurable and reusable internal LLM powered RAG framework

- Engineered a robust Face Recognition framework integrated into HAJIRA's Automated Facial Attendance System, used by multiple companies.- Optimized Facial Recognition for 3x faster face detection and recognition on edge devices from video/live streams.- Reduced inference time by 50% through optimization techniques and efficient NCNN implementation for DL models.- Engineered an optimization framework enabling deployment of pre-trained models on low-powered devices via model simplification and pruning.- Orchestrated seamless deployment of pretrained models on resource-efficient devices using cutting-edge MLOps practices.

- Elevated Face Liveliness detection to an impressive 99% accuracy on our proprietary dataset. Integrated this achievement into GazePass, a cutting-edge Passwordless Authentication System.- Constructed a synthetic NID image dataset using conditional GANs, facilitating rigorous model evaluation on regional datasets.- Reduce the model size by exploring the Model optimization strategies such as Knowledge Distillation without loss of validity for the Face Recognition pipeline.

- Instituted an efficient data processing pipeline to scrape, organize, and clean data extracted from thousands of online sources to build a Bangla extensive textual corpus for the bunch of Bangla NLP tasks.- Built a PyTorch-based fully-fledged Bangla NLP library for a bunch of NLP tasks such as Word2Vec, BERT, NER, POS Tagging, etc.- Trained ELMo model, a novel way to represent words in vectors space, using self-developed large Bangla corpus which incorporated in POS tagging task improving accuracy by 2%.