At variA/Bly, I fine-tune and evaluate domain-specific grounding and judge models that check AI-generated answers against source documents. I benchmark their accuracy, false-positive rate, missed errors, and latency, then calibrate scoring thresholds.
At SpazorLabs, I contribute to production-oriented LLM pipelines and AI agent systems. My work includes agent orchestration, state management, structured outputs, and evaluation for reliability.
Through my independent Modus_X research, I explore memory-efficient recurrent architectures as alternatives to attention-based context storage. I design controlled comparisons with Transformer baselines across memory retention, associative recall, language-model quality, and inference memory.
My projects include a reproducible benchmark comparing GRU and Transformer models on associative recall, and work on calibration and selective prediction for financial QA. I also fine-tuned Qwen2.5-3B-Instruct with QLoRA and built a reproducible training and evaluation pipeline.

