Designed RAG pipelines using LangChain and LlamaIndex with semantic chunking and metadata filters for legal Q&A.

Deployed autonomous LLM agents with tool use, memory, and API integrations for real-time multi-modal task execution, including automated enterprise email workflows.

Applied parameter-efficient fine-tuning (LoRA/QLoRA), optimized deployments with quantized models, and implemented LangSmith-based tracing and evaluation pipelines for observability.

Trained a 110M-parameter BERT model for Nepali and created the first Nepali NLP benchmark suite (text classification, NER, sentiment analysis).

Developed an Active Learning framework with entropy, margin, and least-confidence sampling, achieving near SOTA results with minimal annotation.

Built scalable data pipelines and retraining loops, extending support to computer vision tasks like image classification and object detection.