>_ mohar@portfolio:~/experiments$
experiments.
Hands-on evaluations, stress tests, and reimplementations of AI systems — taking ideas from papers and turning them into something I can break.
Legal Summarization EN → HI
Comparing zero-shot, few-shot, and chain-of-thought prompting for English→Hindi legal text summarization on the MILDSum dataset using Qwen3.5-4B (Q4_K_M).
LLMQwen3.5-4BPromptingHindiEvaluation
Paragraphs to Atomic Facts
Splitting legal-news paragraphs into atomic legal propositions — a three-stage LLM pipeline, a merged check/split/fill loop, and a stage-1-only version that does it in one call, each with a fact-by-fact manual check of what the model got wrong.
Legal NLPgemma-4-31bPipelineAblation