Peijin Li

I’m an AI Product Engineer who spots the real bottleneck, makes the architecture and scope calls, and ships production systems from 0→1. I’ve built end-to-end AI platforms—autonomous agents, production RAG, and multi-agent workflows—inside fast-moving AI teams where ambiguity is the default.

At xDan AI, I identified when off-the-shelf LLM Q&A couldn’t handle domain-specific financial queries, then evaluated retrieval vs. fine-tuning vs. prompt engineering and led the production RAG work that became the default answer path. I also recognized manual annotation as the true bottleneck and built an automated synthetic data generation pipeline—producing 100K+ instructions—to accelerate improvement.

I’ve owned evaluation and deployment decisions as well: I designed an end-to-end LLM evaluation framework using MT-Bench and C-Eval, enabling stakeholders to benchmark 20+ candidate models before go-live and cutting evaluation cycle time by 60%. I partnered with research to turn multi-agent orchestration (task planning, tool integration, memory, and execution replay) into a production system robust enough to run daily.

Earlier, I built and shipped data products and prediction systems—from AWS pipelines for on-demand access to UN Comtrade data, to manufacturing risk prediction using NLP plus XGBoost/Random Forest ensembles (55% to 73%). I also founded an autonomous recruiting and interview scheduling agent that coordinated sourcing, evaluation, scheduling, interviews, and hiring decisions end-to-end in 30 days—backed by resume parsing, candidate assessment, secure storage, and recruiter collaboration workflows.