Work Experience

Agent engineering, FDE delivery, and production hardening.

Orient Feeling (Hangzhou, China)

2026.9 - Present

AI Agent Engineer

Owned the agent runtime architecture for a sleep-health IoT product from day one: conversational workflows, task orchestration, device control, persistence, and memory.

  • Unified analysis, reporting, and deterministic business commands into a single conversational workflow architecture with resumable execution, consolidating previously fragmented pipelines.
  • Designed a versioned task lifecycle and a two-phase device-control protocol (prepare, confirm, execute): hardware actions stay user-confirmed, conflict-checked, and recoverable when a device call fails mid-flight.
  • Closed the end-to-end hardware control loop through a public Task API (intent validation, device-capability checks, bounded outbound calls with rate-limit classification), letting the agent safely operate physical sleep devices.
  • Built the persistence and long-term memory layer (task state, control plans, audit evidence, conversation memory) that survives restarts, with memory reads that degrade gracefully instead of blocking users.
  • Delivered incremental answer streaming with an explicit failure contract, so chat responses appear as they are produced instead of in one blocking wait.
  • Hardening the three production analysis flows (health advice, sleep recommendation, and metric advice) for real health-database and LLM integration, grounding every output in cited evidence.

Baidu AI Cloud (Beijing, China)

2026.7 - 2026.9

Medical Industry Strategy FDE

Owned industry strategy, the 0-to-1 customer master-data model, and presales analytics for Baidu AI Cloud's medical vertical.

Industry strategy

  • Designed the interview framework and led 10+ expert interviews to rebuild the medical vertical's strategic read, decomposing leading cloud vendors' revenue mix, customer segments, product lines, and revenue-recognition logic.
  • Cross-validated internal operating data to set the vertical's customer-segment, product, and resource priorities across three battlegrounds: hospitals, public-health reform, and pharma/medtech/biotech.

Data engineering

  • Designed the 0-to-1 customer master-data model for pharma/medtech/biotech, abstracting legal entities, groups, R&D pipelines, financing/BD, tenders, and sales status into governed core objects.
  • Defined uniqueness, relationship, and history-traceability rules governing multi-source name collisions, duplicates, and entity conflation.
  • Orchestrated a multi-agent pipeline covering multi-source collection, entity resolution, relationship recognition, evidence verification, and incremental updates, with quality gates at ingestion.

Presales enablement

  • Connected master data to the internal CRM, fusing R&D, procurement, competitive-migration, and sales-follow-up signals into a single account view.
  • Built customer-value, demand-window, and opportunity-status scoring calibrated by CRM feedback, distilling a priority-account pool and a sales battle dashboard.
  • Mapped high-potential accounts' budget sources, decision chains, and competitor footprints against Baidu's compute, data-platform, and MaaS offerings, producing battle cards and visit agendas that turned industry insight into customer meetings and pipeline progress.

Qianshu Wenyi Technology (Canton, China)

2026.5 - 2026.7

Medical AI FDE Engineer

Medical record QC solution

  • Interviewed 10+ hospital stakeholders and converted terminal medical-record QC workflows into product fields, status models, and product paths.
  • Modeled case status, defect adjudication, rule evidence, scoring, responsible doctor, rectification deadline, review trace, and report output.
  • Delivered page structure, field model, product spec, and development plan for presales product review and R&D alignment.

AI medical record QC product

  • Built and deployed a browser-based AI QC product covering AI screening, manual review, doctor rectification, recovery review, score recalculation, and final release.
  • Linked each defect to evidence, rule clause, deduction standard, AI judgment, human decision, and operation logs.
  • Connected reviewer actions to score changes, A/B/C grading, downgrade defects, and report templates. View Product

Canchen Technology (Shenzhen, China)

2026.1 - 2026.5

Multi-Agent Product Developer

Rare Disease Agent

  • Designed tool routing for case parsing, HPO extraction, ontology retrieval, candidate ranking, evidence aggregation, and structured reporting.
  • Planned evidence linking across HPO, OMIM, Orphanet, and PubMed, connecting phenotype, disease, gene, and literature signals.
  • Defined state objects for case_summary, phenotype_package, candidate_list, evidence_bundle, audit_notes, and final_report.
  • Built an agent evaluation loop covering HPO extraction accuracy, candidate disease recall, evidence citation quality, tool-call success rate, and report iteration.

MedrixAI Scientist multi-agent research platform

  • Contributed PaperAgent and GraphAgent workflows for literature review, scientific writing, figure generation, knowledge-base Q&A, and clinical case analysis.
  • Built Chat + Workspace + Right Panel interaction patterns with Reasoning, References, Results, and Files surfaces.
  • Planned knowledge-base reuse for PDFs, meeting recordings, Feishu docs, chat records, and research files. Open Workspace

Yulingzhikong (Canton, China)

2025.6 - 2025.12

Co-founder · Commercialization

  • Led 30+ B-side field interviews and repositioned an emotional companion robot from broad consumer use to higher-value eldercare and special-education scenarios.
  • Owned seed-round materials including market analysis, business model, financial forecast, and business plan; helped the team secure RMB 4M angel funding.
  • Modeled BOM and service marginal cost with hardware suppliers, then built hardware-plus-subscription pricing tiers based on willingness-to-pay research.
  • Built channel relationships with eldercare and rehabilitation institutions, supporting device pre-deployment, user-data collection, and product iteration.

HiDream.ai (Shanghai, China)

2024.6 - 2024.9

AI E-commerce Product Intern

  • Led multimodal generation quality evaluation for e-commerce image and video scenarios, defining bad-case taxonomy across lighting, white balance, composition, subject consistency, text alignment, rhythm, and frame consistency.
  • Owned training and evaluation sample deduplication, denoising, and quality filtering; built data-selection rules for product images, marketing videos, and localized materials.
  • Contributed prompt templates, CFG tuning, negative constraints, and style-control strategies for product display, model shots, scene images, and short-video covers.
  • Defined Southeast Asia localization weights for color, clothing, and scene-style terms, reducing underexposure, cultural mismatch, oversharpening, and subject-style drift.
  • Converted behavior tracking and bad-case analysis into PRDs for local repainting, style transfer, and image cleanup. Creator DAU increased by 19k+ after launch; browsing time improved 25%, monthly retention improved 38.8%.

ByteDance (Shanghai, China)

2023.12 - 2024.6

Strategy Product Intern · Data Analyst

Strategy product

  • Built tiered SKU portfolio strategies for 3C categories, balancing traffic SKUs, hero SKUs, profit SKUs, gifts, and installment incentives. Brand daily ROI increased 76%.
  • Executed back-to-school and 618 campaigns for 38 merchants; used hourly GPM racing to recover traffic under competitor price pressure. Conversion increased 35%, brand sales grew 212%.
  • Cleaned after-sales data and extracted high-frequency pain points such as OCR recognition and keyboard switch feel, feeding product-page visualization and live product SOP improvements.

Data analysis

  • Decomposed B-side merchant profit structures and promoted channel-exclusive SKUs to isolate price comparison. High fee-rate exposure decreased 60% year over year.
  • Reweighted ad allocation by inventory depth, delivery reliability, and historical ROI, improving traffic efficiency under budget constraints.
  • Built Aeolus / Power BI real-time dashboards for inventory and price-risk monitoring during Double 12, supporting timely stock-lock, delisting, and campaign-intensity adjustments. Related brand GMV grew 413% year over year.