Yuhan Cheng

University of Toronto

University of Toronto

Chinasoft International

• Working in the AI recruitment domain — candidate parsing, job-requirement understanding, candidate-to-role matching, multilingual content generation. Built an in-house Multi-AI-Agents orchestration framework (no LangGraph) driven by events, wiring an end-to-end pipeline: Clarify → Requirement → Entity Parsing → Matching → Validation → Auto-Correction → HITL review.
• Led the migration of the event bus from Kafka to Inngest with dual-track compatibility, and refactored the backend from Python to TypeScript (Node + Next.js); built an observability panel for the previously opaque async agent chain — agent state, rule execution, logs, and version rollback.
• Addressed noisy, conflict-prone business data with an in-house Semantic / Logic / Syntax 3-layer validation system and an Auto-Correction repair loop, combined with reverse attribution (probabilistic localization across DataObject / Rule / Relation) for iterative refinement — materially raising first-pass quality.
• Applied Context Engineering and Progressive Disclosure to design Agent Skills / Tools selection; distilled internal business knowledge into Claude Code-embedded Skills and an Obsidian Wiki graph knowledge base, loaded on demand to cut prompt noise and token cost.
• Studied and applied state-of-the-art agent design principles from Wiki LLM, Hermes Agents, Harness Engineering, and OpenClaw; leveraged Claude Code / Gemini / Codex for Vibe Coding to compress 1–2 week features and refactors into a few days.

DMD Building Systems

• Led the creation of an LLM-powered DXF drawing parsing and classification platform that automated annotation and archiving of CAD drawings, replacing slow manual workflows in the firm's BIM/CAD pipeline.
• Built a Python + LLM pipeline: DXF parsing to block exploding, parent-child traversal to record chain information, semantic classification to identify wall / door / window / text-annotation categories.
• Used the ezdxf library to build an ancestor-descendant dictionary for GPT/Gemini context, added parallel caching and incremental updates — tripled parsing speed.
• Delivered JSON outputs and visual annotations automatically written back to the design-management platform, closing the loop from analysis to actionable design data.
• Researched and synthesised the latest AI-for-AEC (Architecture, Engineering & Construction) literature, designing end-to-end system architectures and delivering technical reports.