Things I've built — coursework, research, and side projects.
Post-training a Dual-Control Tool Agent
2026pass^1 11.2% → 86.4%, pass^4 0% → 66.7% over the base model
SFT + GRPO post-training of Qwen3-8B (LoRA, one A800) on τ²-bench telecom, where the agent must talk a simulated customer through fixes on a phone it cannot touch. Fully programmatic reward, no LLM judge; includes a measured negative result on turn-level credit assignment.
WebAgent
202687% success on 45 held-out WebArena Verified tasks (baseline 67%), $0.006 per task
Browser agent that runs an e-commerce admin console from plain-language tasks, written without an agent framework over the Chrome DevTools Protocol. A cache-first context pipeline cut tokens per task by 91%, and a skill library promotes code only after it replays correctly on the live site.
Multimodal Document RAG
2026+0.171 NDCG@10 and +0.094 answer score over text RAG on held-out ViDoRe v3 (p = 0.0001)
Question answering over PDF report pages. ColQwen2.5 page-image retrieval is fused with dense text and BM25, a retrieval agent picks the evidence pages, a VLM answers, and code verifies every cited quote.
OfferCome
2026Live product, self-hostable with Docker
Open-source AI workspace for job applications, resumes and mock interviews. The AI interviewer asks follow-up questions based on your resume and your previous answers, then writes a review report and a capability profile.
Agent Codebase Analysis Skill
2026Agent Skill for Codex and Claude Code that explains an LLM-agent codebase as a runtime system (harness, context flow, tools, memory, multi-agent coordination) in one HTML architecture report.
yuecao.dev
This website. Static site built with Astro, deployed on Cloudflare Pages.