Data Scientist with expertise in LLMs, NLP, Search, Embedding, and AI systems.
Building and optimizing large language model pipelines — training data, constrained inference, NLP at scale.
Currently at Draup, powering sales and talent intelligence with ML.
LLMs Fine-tuning Constrained Decoding NER RLHF Embeddings PyTorch Hugging Face
constrained-decoding — Forces open-weight LLMs to output exactly one of your taxonomy labels using a trie. No prompt tricks, no post-processing regex — structural guarantee at the token level. Handles multi-label + repeat prevention.
LLM-NER-Offset-Generator — NER training data generation using LLM tool-calling. Gets start_offset / end_offset right, which is the part most LLM-based annotation pipelines get wrong.
annotated-research-papers — NLP/LLM papers I've read, with notes. An archive rather than a live project.
html_tag_annotator — Chrome extension + ML tool to create training datasets fast — tag HTML elements directly in the browser.
web-mark — Chrome extension to highlight and annotate web pages, with optional Google Sheets backup.
Stack: Python · PyTorch · Hugging Face · spaCy · LangChain · FastAPI · LM Studio
Focus areas: LLM fine-tuning · constrained decoding · NER · embeddings · RLHF
sachinkalsi.github.io · Blog · LinkedIn · @sachin_kalsi · sachinkalsi15@gmail.com