AI applications
Structured understanding of documents, papers, and long-form material — RAG / OCR in practice.
About
About me
I like turning ambiguous problems into executable systems, then using documentation, tests, and evidence to explain what the system really does.
01CS / AI toolingPROGRAMS
02Open sourceOPEN SOURCE
03Evidence-ledEVIDENCE
04Limits shownLIMITS
01 / FOCUS AREAS
Focus
Capability groups backed by public projects, not a wall of badges.
Agent tooling
Orchestration and workflow engineering with Skills, MCP, and subagents.
Testing engineering
Case generation, evaluation, and self-healing loops that make systems verifiable.
02 / PRINCIPLES
Engineering principles
Standards applied to code, docs, and this site alike.
- 01Layer the factsSeparate verified facts, inference, and unknown state.
- 02Evidence at the entrancePut quick-start steps, evidence, and limitations at the project entrance.
- 03No inflated claimsDo not describe static checks as live-service success, or forks as original work.
- 04Credentials out of artifactsKeep credentials out of code, logs, repositories, build artifacts, and public profiles.
03 / RESUME
Resume
CS student working on AI applications, agent tooling, and test engineering, focused on RAG, OCR, and document understanding — with hands-on experience shipping AI test agents from case generation to self-healing loops.
Projects
Tencent Maps UI Automation Testing Agent (framework team contributor); RAG-based Pharma Data Analysis Q&A Assistant (key contributor); Intelligent Continuous Testing Platform (solo design & development).
Academic highlights
Silver in "Internet+" provincial competition, university second prize, provincial third prize in the Chongqing AI model competition; SCI Q1 paper under review.
Skills
Python / Pytest / Playwright / FastAPI / MySQL / Docker / Git / Linux; agent mechanics and evaluation methodology.
Details
Projects
Tencent Maps UI Automation Testing Agent
Framework team contributor2026.06 – 2026.09- Agent orchestration & case generation: built an LLM-agent quality & efficiency system layered into five roles orchestrating 13 Skills and 3 in-house MCP services, closing the "natural-language requirement → automated case → knowledge retention" loop; linked TAPD/Figma to auto-generate executable case.py from requirements, ~10–15 min per case versus 1–2 hours handwritten, covering 17 business domains and 600+ generated code units.
- MCP tooling & multimodal perception: built device-control and online-diagnosis MCP tools, integrating the Tencent Maps MCP to resolve coordinates and construct test data; fused UIAutomator, OCR, OmniParser, and OpenCV with chained fallback to locate self-drawn widgets lacking resource-id; scrcpy long-lived projection cut screenshot latency from ~1s to 30–100ms.
- Knowledge engineering & semantic retrieval: designed a file-based knowledge graph with same-source static reasoning over UI indexes from 23 developer repos so the agent reads dev code before generating; AST three-stage filtering over 1500+ private helpers, with jieba + BM25 semantic retrieval to reuse common methods and avoid reinventing wheels at scale.
- Reliability & self-healing loop: guarded generation with intent locking, gated exploration, and nine-dimension self-review; enforced Lint gates and step-wise assertions for stable code, pruned flaky cases via health backtest pass rates, and auto-regenerated patches from evidence packages into a self-heal loop.
RAG-based Pharma Data Analysis Q&A Assistant
Key contributor2025.10 – 2026.02- Goal: built an LLM data-analysis Q&A assistant over full-chain pharma data (procurement, retail, inventory) and 800+ business tables on Dify, LangChain, Milvus, and MySQL, turning sales questions into executable SQL and readable insights.
- Application architecture: owned the core LLM chain design, workflow orchestration, and model integration on Dify, shipping "question understanding – schema retrieval – SQL generation – result interpretation – human feedback" to lower the barrier of understanding complex table schemas.
- Schema understanding & field recall: enriched the data dictionary with metadata, vectorizing table names, field names, meanings, and common enum values as rich text; combined business-rule filtering, question-type classification, sample pairs, and CoT to steer the model toward relevant tables and fields, stabilizing SQL generation.
- Evaluation & self-correction loop: added SQL pre-execution with self-correction, feeding MySQL errors back to the LLM for rewrite; built ~200 Golden Dataset entries and assessed retrieval/SQL quality with RAGAS metrics, execution accuracy, and Bad Case feedback — execution success rate up from ~70% to 90%+.
Intelligent Continuous Testing Platform
Solo design & development2025.10 – 2026.02- Platform & framework design: built a five-layer API automation framework with Python/Pytest/Requests, auto-generated contract cases from OpenAPI validating status codes, response schemas, and business error codes; assembled a lightweight platform on Vue.js + FastAPI + MySQL + Redis + Docker to manage projects and cases.
- UI automation & performance testing: implemented Playwright PO-pattern business cases with Trace, screenshots, and recording for replay; designed ramp-up, stability, and peak load tests with JMeter/Locust, reporting QPS, P95, and error rates while locating slow endpoints.
- AI enhancement & quality loop: wrapped Pytest/Playwright/Locust execution as async Redis-queued platform tasks; introduced an LLM case-generation pipeline (draft → human review → Bad Case retention); report pages aggregate pass rates, failure reasons, and trends into a closed loop.
Academic highlights
- Competitions: Silver, 2025 "Internet+" Innovation & Entrepreneurship Competition; university second prize, China International College Students' Innovation Competition; provincial third prize, Chongqing AI Large Model Innovation Application Competition.
- Paper: "LHAF: A Lightweight Hierarchical Adaptive Framework for LLM-based Multimodal Emotion Recognition in Conversations" (submitted to SCI Q1).
Skills
- Test development: equivalence partitioning, boundary value, scenario, state-machine, and decision-table design; full test lifecycle from requirement review through regression; TAPD and internal case/defect platforms.
- Engineering & tooling: API automation with Python/Pytest/Requests; UI/mobile automation with Playwright, Selenium, uiautomator2, and ADB; load testing with JMeter and Locust; Linux, FastAPI, MySQL, Docker, Git.
- AI application & testing: understands RAG, Harness, and Memory agent mechanics; fluent in MCP, Skill, and Subagent frameworks; builds Golden Dataset and Bad Case feedback loops with hallucination, refusal, formatting, and multi-turn regression suites validated by keyword, JSON Schema, similarity, and rule assertions; production experience in auto-updating agent cases with human review.
- Identity
- CS student / AI tooling builder
- Location
- Chongqing, China
- Focus
- Projects · Articles · Notes
- Resume
- resume.md ↓
- Positioning
- CS student building AI applications, agent tools, and test engineering, with a focus on RAG, OCR, and document understanding.
- 2508807574@qq.com
- Entry
- GitHub Profile ↗