york Zhou

Backed by Baidu Ventures, Sinovation Ventures, Source Code Capital, and Zhongguancun Venture Capital Building AI-native products at a late Series A startup, with a core focus on pioneering agentic workflows and Agent-to-Agent (A2A) ecosystems in the recruitment industry. 🚀 A2A Platform — MoltOffer (moltoffer.ai) • Driving the product design of an Agent-to-Agent recruitment network currently hosting 1,100+ autonomous agents across both candidate and recruiter ecosystems. • Shaping an open agent ecosystem by defining OpenAPI strategies, enabling external developers to build and deploy custom agents on the platform. • Core focus areas: Agent identity, capability discovery, interaction protocols, and cross-agent trust mechanisms. 🤝 Headhunter Agent Collaboration Network (OpenClaw-based) • Designing a collaborative agent network where agents representing different recruiters discover, connect, and cooperate seamlessly across candidate sourcing and placement workflows. • Defining agent capability boundaries, interaction flows, and trust mechanisms tailored to ensure accessibility for non-technical users. ⚙️ Internal Agent-Friendly Transformation • Spearheading the refactoring of internal systems to be fully agent-readable and operable. • Exposing core interfaces and data layers to empower autonomous agents to act reliably and efficiently within the organization's workflows.

Led the 0-to-1 development of a Few-Shot self-service module for TikTok's cross-border content governance platform, drastically accelerating model iteration cycles from 3–4 weeks to just 3.4 days. 🔍 User Research & Workflow Optimization • Conducted deep-dive interviews with ops stakeholders to map workflows, successfully identifying and addressing 3 core bottlenecks: low risk recall, lack of ML literacy among operators, and cold-start data challenges. ⚙️ Model Architecture & Strategy • Evaluated 6+ model architectures and drove the adoption of a hybrid small-model + fine-tuned LLM approach. • Achieved robust metrics (R86/P72) with a minimal 50-sample cold start and 4-hour training turnaround. 🤖 Automated Prompt Engineering (AutoPE) • Designed an AutoPE pipeline and standardized 4 reusable prompt templates, slashing prompt engineering time from 90 minutes down to 15 minutes. 📊 Data & Sample Management System • Architected a highly efficient sample management system from scratch, reducing data contamination rates from 20% to 2% and cutting collection time by 60%. 🏆 Key Impact / Results • Successfully shipped 25 custom models. • Accelerated the go-live cycle by ~85% (from 3–4 weeks to 3.4 days). • Boosted incremental risk recall by +8.7%.

🤖 LLM Data Classification Agent (Python) • Pipeline Scaling: Architected an end-to-end data pipeline that scaled daily processing throughput by 100x (from 1K to 100K), reducing manual operational effort by 90%. • Prompt Engineering: Iterated prompts across 10 rounds leveraging the CRAP framework, dramatically lifting the classification F1 score from 40% to 95%. • Product Optimization: Identified 6 high-friction conversation categories through data analysis. Leveraged these outputs to guide model fine-tuning, successfully cutting the user thumbs-down (negative feedback) rate by 27%. 📊 Model Evaluation & Competitive Intelligence • Eval Framework Design: Co-designed a 3-dimensional evaluation framework benchmarking against 5 key industry competitors. • Data Quality: Produced 1,200+ highly calibrated labeled samples, improving BadCase detection efficiency by 30%. • Strategic Impact: Built an API competitive matrix and delivered 4 comprehensive market reports, directly informing and shaping the product's differentiated pricing strategy.