Varun Thanneeru

Data Scientist

Turing Intelligence • Temps plein

Jun 2025 - Jun 2026 • 1 yr

•Fine-tuned LLMs on curated Verified Instruction Following (VIF) datasets using LoRA, reducing GPU memory requirements by ~4x during training; co-developed a FastAPI-based evaluation service for automated metrics (exactmatch, rubric-based) to benchmark instruction-following fidelity. • Built deterministic tau-bench-style environments for agent tasks with simulated user/tool/database scenarios and expected action-output traces to objectively measure LLM agent task completion. • Collaborated on a pipeline to generate multi-turn Agent Task Completion dialogues enforcing correct tool-call sequencing, multi-step reasoning, and graceful failure handling for downstream SFT training data. • Contributed to and evaluated Computer-Use Agent (CUA) pipelines using the OSWorld benchmark, measuring screengrounding accuracy and multi-step task completion across 200+ tasks and 5+ environments.