Master's student in Cybersecurity at NTU Singapore, working at the intersection of AI, security, and product design. I translate technical capabilities into practical, user-oriented solutions — LLM applications, RAG architectures, and backend systems that emphasize reliability, scalability, and data privacy in sensitive domains.
About
Experience
Experience
AI Product Associate Intern
Shipped an LLM-as-a-judge evaluation pipeline (TTS synthesis → blind multi-model transcription → 3-vote majority) that lifted a voice agent's local place-name pronunciation accuracy from 78% to 93.8% across a 349-entry dictionary, packaged as a reusable backend component with edit-safe bulk correction. Drove 0→1 feasibility and architecture for an LLM-tool capability letting voice agents traverse phone menus (DTMF/IVR navigation), hardened by an 8-agent adversarial code review that caught 4 production-blocking defects. Delivered client- and event-facing products end to end: a PII-hardened QR lead-capture web app for a live product event, a single-file visual mockup for a partner client demo on deadline, and requirement-to-voice-stack mapping for a competitive government AI tender. Authored ~50 product specs, feasibility studies, and PRDs on a production voice AI platform, turning ambiguity into documented, decision-ready scope.
Researcher (Part-time)
Built a RAG-based internal Q&A system over policy and process documents. KNN retrieval on long-document corpora; scenario-grounded evaluation reaching 96% Recall@N and 4.4/5 answer relevance.
Researcher (Part-time)
Multi-agent reinforcement learning (A2C-based) for large-scale microservice scheduling. Validated on real-world ByteDance and Alibaba traces; co-authored publication in Expert Systems with Applications.
Product Manager Intern
Led a 6-person cross-functional team to design and deliver a warehouse inventory management prototype through to enterprise-mentor acceptance, owning requirements, prioritisation, task allocation, and testing. Standardised inbound, outbound, inventory-query, stock-alert, reconciliation, and discrepancy-handling workflows, and personally implemented 4 core frontend, backend, and AI features — authentication, role-based access control, workflow configuration, and LLM-powered daily inventory reports.
Education
Nanyang Technological University
Singapore
Xi'an Jiaotong University
Xi'an, China
Publications
Ma, N., Tang, A., Xiong, Z., & Jiang, F. (2025). A deep multi-agent reinforcement learning approach for the micro-service migration problem with affinity in the cloud. Expert Systems with Applications, 273, 126856.
Selected Work
01
MindGap · Claude Code Plugin
A Claude Code plugin that measures where each turn's wall time actually goes — Claude wait, automated tools, or user-gated tools. Zero-dependency Node.js hooks log every UserPromptSubmit / PreToolUse / PostToolUse / Stop to JSONL; a paired skill aggregates per-turn timings on demand. Published to the Claude Code marketplace under MIT.
02
AI Data Visualization Workflow Tool
End-to-end LLM workflow that unifies parsing, cleaning, analysis, and visualization across heterogeneous data (PDF / Excel / CSV). Compresses multi-hour analyst work to ~10 minutes; deployed as a Lark bot for zero-setup access. Built with Coze.
Skill Index
- 01 Product Management
- 02 LLM
- 03 RAG
- 04 AI Agent
- 05 Prompt Engineering
- 06 Workflow Design
- 07 Coze · FastGPT
- 08 TypeScript
- 09 Java
- 10 Spring Boot
- 11 PostgreSQL
- 12 MongoDB
- 13 Cybersecurity
hover to reveal connections