Hanxiao Xiang
Profile
Master's candidate in Industrial Engineering and Management at South China University of Technology, with experience across AI product design, agent and skill engineering, AI voice interaction, technical operations, user research, and data analytics. Capable of owning the lifecycle from discovery, PRDs and prototypes to MVP validation and production delivery, while working deeply with ASR–LLM–TTS pipelines, state machines, workflow orchestration, model fine-tuning, and evaluation.
Education
South China University of Technology (Project 985) · MEng, Industrial Engineering and Management
Donghua University (Project 211) · BSc, Information Management and Information Systems
Experience
INTSIG Information
GEO digital employee: workflow design, long-running execution, and reporting
Mapped the GEO business lifecycle from research and content decomposition/distribution to performance monitoring, defining task parsing, multi-platform collection, AI evaluation, and report delivery. Packaged data collection, KPI analysis, and report generation as callable production-grade skills.
Addressed 6–8-hour multi-assistant collection jobs that timed out and restarted from scratch, platform anti-scraping constraints, and blocked or unanswered WeCom bot messages. Split continuous execution into 30-minute tasks and designed Agent Harness mechanisms including a state machine, task sharding, persistent state, checkpoint resume, failure recovery, and anti-scraping measures.
Separated deterministic execution from semantic decisions: scripts handled crawling, state-machine transitions, and scheduled triggers, while AI handled semantically dependent task routing, interpretation, and data labeling. For reports covering brand mention rate, first-recommendation rate, and brand mindshare, authored content-format, theme-design, and data-citation standards referenced from
SKILL.md; defined agent roles, capability boundaries, and self-introduction behavior inAGENTS.md.Enabled stable long-running execution and result delivery while allowing users to ask questions, check progress, and receive milestone notifications during a task, improving the interaction experience.
Risk-control digital employee: scenario design and policy-monitoring evaluation
Defined business rules, inputs/outputs, and exception handling across six security-team scenarios, designing a four-skill agent chain spanning information retrieval, risk identification, data analysis, and result delivery.
Addressed policy-monitoring false positives by building an evaluation set from historical data, performing error attribution, and iterating prompts, skills, and workflows; reduced false positives from 12 per day to two or fewer.
Internal business-data access for digital employees
Addressed an agent sandbox that supported only HTTP and could not connect to the company's cloud database. Mapped the data-access path and required platform changes, aligned the technical approach and schedule with platform developers and operations, and continued driving prioritization until digital employees could access internal business data.
Binglan Intelligent Technology · AI Customer Service / Outbound Calling
AI customer-service conversation evaluation and improvement
Addressed cases where users recognized the bot or found its responses unnatural in an ASR–LLM–TTS customer-service and outbound-calling product. Reviewed real call cases and attributed failures to ASR errors, model latency, wording, or tone; identified issues with conversational fillers, expressions, and abnormal model output, then configured replacement terms and post-processing rules to improve conversational naturalness and a more human-like experience.
Multi-tenant outbound-calling console
Addressed error-prone cross-system lead imports by independently producing the console prototype and four PRDs with Codex and Figma MCP; defined tenant isolation, status transitions, and delayed initiation. After launch, lead-entry errors fell by approximately 75% and total processing time by 60%; the product was adopted by three business lines.
Business-report automation
Decomposed repetitive data collection and analysis tasks, iteratively built utility scripts through vibe coding, packaged them as callable AI skills, and integrated them into the team's agent—cutting manual report processing from over one hour to three minutes.
XPENG AEROHT
Pre-launch user-requirements research
Conducted pre-launch user interviews, translated ambiguous needs into 50+ hardware, marketing, and service-ecosystem requirements, and produced a cross-functional insight report; designed follow-up research and interview guides for unresolved findings.
Competitive landscape and product research
Analyzed specifications, R&D progress, and resource readiness across 63 competitors in five domestic and international categories through desk research and expert interviews; delivered a tiered competitive landscape report.
AI-tool evaluation and guides for insight work
Evaluated 12 AI products across eight dimensions, mapped high-frequency department SOPs, and led three guides covering DeepSeek, ChatGPT, Nano Banana Pro, reusable use cases, and prompt templates for insight work.
Vertical use-case and market-opportunity research
Assessed 12 major vertical application scenarios against product capabilities, produced one market-opportunity report, and designed three interview plans for distinct expert groups.
Shanghai Rongcheng CPA
Financial-statement checking workflows and deposit-certificate automation
Worked with the business owner on an initial "financial-statement checking agent" concept, using multiple discovery discussions to decompose the checking process, rule boundaries, and accuracy and stability requirements. Identified determinism and controllability risks in an autonomous-agent approach, compared agent and workflow architectures, and drove adoption of a controllable n8n and multimodal-LLM workflow.
Deployed n8n locally, integrated a multimodal-LLM API, and re-engineered SOPs for financial-statement checks and deposit-certificate recognition into a 20+ node workflow. Constrained AI output through a strict JSON-format validator and feedback mechanism; implemented field parsing, rule validation, and exception branches; and exposed Python scripts as local API services for workflow calls, reducing processing time from three minutes to seven seconds—approximately a 30× efficiency gain.
Multi-type invoice-recognition agent
Owned discovery, product planning, technical selection, UI design, and launch of a ReAct-based agent supporting five invoice types; designed tool-use logic and achieved 90%+ end-to-end recognition accuracy.
Non-standard balance-sheet normalization: post-training optimization
Under hard constraints requiring private deployment with limited resources, experiments showed that smaller models struggled with complex normalization rules, particularly debit/credit directions for different accounting categories. This caused numerical errors and noncompliant table structures, with task accuracy below 10%.
Designed synthetic samples targeting these difficult representative cases and built 11,000+ training samples. Validated data formats and accounting-rule logic, filtering out noncompliant samples to keep erroneous data out of training.
Designed the post-training plan with reference to the financial reasoning model Fin-R1 and the Qwen2.5-1M technical report. Used SFT to learn reasoning and output patterns from examples, then applied GRPO reinforcement-learning fine-tuning with three reward types covering compliance, format completeness, and logical consistency, aiming to improve rule application and generalization.
Built a six-dimensional evaluation framework focused on data formats and numerical results. Iterated the data and model approach through error attribution, ultimately improving task accuracy to 85%+.
Heterogeneous-table retrieval evaluation
Used controlled experiments and improved Python data-processing algorithms for heterogeneous-table retrieval; iteratively tested prompt hypotheses and delivered six evaluation reports.
Resource preparation for local LLM deployment
Researched resource and server requirements for local LLM deployment, compared market quotations, and supported server procurement negotiations.
HAKUHODO · Shanghai Hakuhodo Advertising
Brand demand and audience insights
Supported beauty-brand projects using Tmall TMIC and Douyin Ocean Engine for demand insights, audience profiling, competitor analysis, and word-cloud visualization.
AI news and tool-sharing program
Tracked frontier AI developments, tested relevant products, documented usage experience, and shared findings across teams.
Curel selling-point and user-demand analysis
Built a selling-point keyword matrix across Tmall and Douyin, analyzing efficacy and ingredients to identify user priorities and product opportunities for Curel.
Marketing AIGC content and corpus support
Used ChatGPT and Midjourney to support ad-copy and sample-image generation and helped organize a marketing-domain corpus.
Kantar China · Qualitative Research Intern
German supplier research and bid comparison
Independently screened German suppliers, communicated requirements and pricing in English, and compared more than three bids.
Consumer-trend and industry research
Collected 50+ core industry reports for a single engagement, integrated multi-dimensional market data, used ChatGPT to explore consumer trends, and delivered three briefs.
Brand qualitative research and user interviews
Supported respondent screening, discussion-guide design, and 20+ online and offline interviews for Master Kong, Li Auto, Riot Games, and WeChat Channels.
Transcript structuring and user-segment migration analysis
Structured hundreds of pages of transcripts with AI tools, modeled user-segment migration, created visualizations, and delivered four insight reports.
Community & Entrepreneurship
Jizhi Liu AI Open-Source Community
AI technical content and multi-channel community operations
Led WeChat, Zhihu, and Xiaohongshu operations for an AI technical-content and developer community. Tracked papers, GitHub projects, models, and developer-tool releases in AI agents and embodied intelligence, selecting explainers and technical topics based on technical value, community interest, and developer needs.
Used engagement data, community management, and audience tagging to analyze article performance and user preferences and iterate editorial strategy. Authored 10+ articles with 60,000+ cumulative views; representative posts reached 16,000 and 6,000+ views.
Operated an account with 35,000+ followers and communities totaling 5,000+ members; the account was previously recognized among China’s Top 30 AI media accounts.
AI content-production workflow
Decomposed the content lifecycle from topic selection to final draft and built an automated workflow in Coze; ran iterative prompt experiments based on user feedback to reduce formulaic AI writing and doubled drafting efficiency; decoupled the mature logic into standard AI Skills / Agents to lower maintenance and adoption costs.
Mobile and web collaboration workspace, in progress
To address fragmented workflows and team handoffs across information collection, content production, and automated formatting, began integrating existing workflows and iterating on a one-stop mobile and web workspace intended to centralize task, content, and process management. The workspace remains under development and has not yet been formally delivered.
New Year AI Tea Party
Organized the online "New Year AI Tea Party," owning promotion, sponsorship communication, speaker outreach, and livestream operations; generated 5,000+ views when the account had just passed 10,000 followers and converted 700+ users into private communities.
Technology partnerships
Led business outreach and secured collaborations with InfoQ, AWS, and other technology organizations.
InternLM LLM Bootcamp
LLM bootcamp: developer support and cohort completion
Supported 3,000+ learners on model deployment, fine-tuning, and OOM issues. Reproduced GPU-memory constraints and coordinated technical troubleshooting, providing actionable solutions such as DeepSpeed ZeRO CPU Offload and reduced KV-cache allocation.
Used Feishu tools and scripts to aggregate assignment status, verify course completion, and distribute certificates, supporting large-scale online interaction and reliable cohort completion.
Dream Workers Program · Social-Impact Venture
Social-impact venture website
Co-founded a public-interest initiative supporting children with Down syndrome and intellectual disabilities, promoting social care, equal treatment, and joyful development while exploring digital channels for broader impact.
Led front-end and IT work, building and maintaining the organization website with HTML, CSS, and JavaScript and using Git for both remote and local code backups.
Social-impact game mini program, planned
Planned an online interactive-game mini program to raise public awareness of children with intellectual disabilities and increase the initiative's visibility. As of Nov 2023, this remained a planned initiative with no evidence of delivery.
OpenMMLab MMPose
DeepFashion2 dataset integration and open-source delivery
Added DeepFashion2 support to MMPose 1.x, independently implementing the dataset class, training configuration, pretrained model, and logs; authored Chinese and English Dataset Zoo instructions to form a reusable training solution.
Responded to maintainer review by removing Chinese comments from configuration files, adding JSON-format training logs and the pretrained model, and syncing the
dev-1.xbranch for CI checks; drove the PR to merge.
Projects
“Transaction” AI Hackathon · Pixel Memory Exchange Game
Intelligent Operations Diagnostics for a Private Healthcare Provider
Data Management Capability Maturity Assessment for Guangzhou Company Y
LLM-Based Business Rule Generation
Inventory and Sales Management System
Product-Impact Study of Hotel Self Check-In Kiosks · Shanghai Innovation Program
Computer Application Competition · Garment Silhouette Recognition
Mathematical Contest in Modeling
Wordle player forecasting and puzzle-difficulty evaluation
Used a limited dataset of 300+ samples and LSTM to forecast daily player counts, then designed VL-LSTM to mitigate underfitting caused by scarce training data, achieving 98.9% model accuracy. Extracted features using linguistic knowledge and built a BP neural network to predict the share of players at each attempt count, achieving 93.3% accuracy.
Combined mathematical modeling, grey relational analysis, and entropy weighting to construct objective and subjective difficulty variables; clustered puzzles by their statistical features with DBSCAN, then trained an SVM on the cluster patterns and coupled it with the forecasting model to evaluate daily puzzle difficulty.
Technical Writing
- Independently authored and maintained a one-stop Linux learning and operations guide based on hands-on learning and deployment experience.
- Published technical articles on CSDN covering OpenMMLab troubleshooting and Chinese documentation for Drools 8.0.