Round 2: 200 → 7 (2nd into final); Round 3: 7 → 3 and hired
Designed a four-stage agent workflow — extract → reuse → draft → fact-check — decomposing long-content knowledge reuse into verifiable tasks.
Owned inter-agent task passing, prompt design, output constraints and workflow orchestration, turning one-shot LLM generation into a multi-stage, decomposable, verifiable pipeline.
Independently delivered requirement definition, PRD, system architecture, a runnable prototype and product demo — the full loop from problem definition to product validation.
Independently produced a 12-page product proposal and pitch deck, and presented in a live demo — able to translate technical solutions into business language.
Horizon Startup 72h Hackathon | OPC Founder AI WorkbenchBackend Lead2026.07
Used Claude Code as the primary coding agent to deliver the backend 0→1 with 15 APIs in ~24 hours.
Built an "AI generate → human review → hand-written tests → bug localization → fix" collaboration loop, catching 3 critical edge cases, including a false trigger of req.on('close').
Wrote HANDOFF.md to structure project background, architecture, APIs and context so a fresh AI agent can quickly resume development.
Established an AI-native collaboration workflow: AI handles high-frequency execution; humans own requirements, architecture judgment, testing and quality control.
Keywords AI Coding · Backend API · Debug · Testing · Code Review
Solved the problem of turning a seedling farm's WeChat chat records into bookkeeping data by designing a full pipeline: natural-language input → AI extraction → structured ledger → Feishu template.
Independently built the AI processing logic, frontend interactions and app, turning unstructured chat into standardized income/expense records with one-click Feishu template generation.
Covered AI information extraction, structured data handling, UX and deployment — a closed loop from real problem discovery to shipped product.
Keywords LLM Application · Information Extraction · Structured Data · Real-world AI
Continuous AI Building | Personal AI Tool MatrixIndependent Developer2026 – Present
Resume Tailor Job Workbench: an all-in-one job-hunting workbench (JD analysis → résumé tailoring → AI mock interview).
Unified Video Tool: video content extraction, text processing and image/PPT generation.
Photo Zine Studio: an AI tool that turns ordinary photos into posters / visual content.
Continuously exploring agents, content generation, multimodal processing and AI product workflows, building a fast idea-to-demo habit.
Round 2: 200 → 7 (2nd into final); Round 3: 7 → 3 and hired
Defined the problem from 0: identified repetitive labor in long-content knowledge reuse, and designed a four-stage AI workflow — extract → reuse → draft → fact-check.
Independently produced requirement definition, PRD, system architecture, a runnable prototype and product demo — the full problem → solution → validation loop.
Independently produced a 12-page product proposal + pitch script, and presented in a live demo, translating technical solutions into business-friendly language.
Owned inter-agent task passing, prompt design and output constraints, ensuring the multi-stage pipeline is decomposable, verifiable and deliverable.
Started from a real inefficiency in the family seedling farm, turning the manual "WeChat records → bookkeeping data" process into an automatable AI workflow.
Designed and shipped the full natural-language input → AI extraction → structured ledger → Feishu template chain, owning all development and deployment.
Deployed, used it for real and iterated continuously, validating AI's real user value in a vertical niche scenario.
Keywords Business Understanding · Process Mapping · AI Deployment · User Value Validation
Used Claude Code as the primary coding agent to deliver the backend 0→1 with 15 APIs in ~24 hours, owning delivery cadence and priorities.
Built an "AI generate → human review → hand-written tests → bug localization → fix" quality loop, catching 3 critical issues, including a false trigger of req.on('close').
Wrote HANDOFF.md to capture project context so a new AI agent or teammate can quickly take over development.
Keywords Project Driving · Delivery Cadence · Quality Control · AI Coding
Continuous AI Building | Personal AI Tool MatrixIndependent Developer2026 – Present
Resume Tailor Job Workbench: an all-in-one job-hunting workbench (JD analysis → résumé tailoring → AI mock interview).
Unified Video Tool: video content extraction, text processing and image/PPT generation.
Photo Zine Studio: an AI tool that turns ordinary photos into posters / visual content.
Continuously validating ideas into runnable demos, building a stable 0→1 development habit.
RK3588 Multi-stream Video AI Inference Pipeline | Embedded AISource Reading & Pipeline Analysis
Systematically read and mapped an RK3588 multi-stream video AI inference project, analyzing the full end-to-end data flow: FFmpeg ingest → MPP decode → NV12/BGR conversion → RKNN NPU inference → post-processing/NMS → multi-model fusion → MPP encode → RTMP push.
Focused on NPU I/O, video frame processing, and decoupling of producer and consumer queues to understand the full embedded AI video inference pipeline.
Familiar with C++ / Linux / FFmpeg / MPP / RKNN / NPU engineering code.
Keywords Embedded AI · C++ · Linux · NPU · Pipeline Analysis
Internship
Smart Agriculture Research Institute | Data Annotation Intern2026.01 – 2026.02
Handled image data annotation and quality checks, completing ~4000 annotation points/day at 95%+ accuracy.
Performed annotation, review, and correction according to project standards, gaining familiarity with AI data production and basic data quality control.
Education
Tianjin Normal UniversitySoftware Engineering | Bachelor2024 – 2028
