Jing Wang

• Motivation: Conventional RAG knowledge bases rely on flat retrieval and struggle to express the logical relationships among rules, corollaries, and exceptions.
• Research focus: Build a logical hierarchy of definitions, axioms, theorems, corollaries, and exception rules — and address how to identify these knowledge types from a continuously growing document corpus, establish their logical relationships, and detect conflicts between new and existing rules.
• Contribution: A deployable framework that helps enterprises, law firms, and regulators build structured knowledge bases. As new documents are continuously ingested, it supports not only retrieval but also rule-conflict detection, distinguishing sanctioned exceptions from substantive contradictions — improving the auditability of knowledge systems in highly regulated domains such as law and finance, and lowering downstream maintenance cost.
• Foundation: built on the llmwiki I developed — a self-maintaining agentic knowledge base — so the auditable axiomatic structure advances on a real, running system rather than in the abstract.

Open-source, cross-platform, cross-agent memory & session-sync for Claude Code & Codex — a single Git repo (using symlinks/junctions) keeps rules, preferences, and accumulated facts in sync across Windows, macOS, Linux, and iPhone, with automatic hooks, a /session-memory skill (save/read/get), and privacy-first redaction.

• Automated pipeline: a loop + AI pipeline auto-curates the day's AI news and papers, generates a video brief, and publishes it to Xiaohongshu.
• Data-driven iteration: it analyzes engagement data, reviews performance, and adjusts topic selection and framing — closing the loop.
• Experiment goal: a controlled testbed comparing human-in-the-loop and fully autonomous workflows, probing the limits of automation and the best points for human intervention.

Master's thesis (full marks):

• Turned unstable absolute sentiment scoring into a stable pairwise-contrastive task.
• Built a 7-group sentiment reference scale (8 intervals) via a single round-robin tournament over a base news corpus; used RAG to retrieve topic-matched anchor news, then binary-search pairwise comparisons to place each target news at one of 8 fine-grained sentiment levels.
• Outperformed FinBERT and FinDPO across multiple backtesting strategies.

Frankfurt, Germany

• Generate regular backtesting and independently analyze red traffic-light breaches for margin parameters.
• Design and document the workflow for Bundesbank Credit Facility reporting.
• Built an agentic, self-maintained knowledge base (llmwiki) and shared AI concepts (agent skills, harness engineering) to raise team AI literacy.
• Initiated a LangGraph + LangChain AI coding agent with a ReAct loop and tool use, built on the Databricks LLM API to safely interact with confidential data and support custom agents/workflows.

Remote

• Built and launched a WeChat mini program for internal factory collaboration end-to-end — requirements, front-end & back-end development, cloud deployment, and post-launch iteration — using Claude Code to accelerate delivery.
• Deployed on Alibaba Cloud for a 100+ employee factory; owned database design, access control, API development, and maintenance.

M.Sc. in Applied Data Science · Frankfurt am Main, Germany

• Specialization in Data Analytics, Statistical Modeling, Machine Learning, and Finance.
• Master's thesis (grade 1.0): Leveraging Large Language Models for Sentiment Analysis in Portfolio Management — developed a RAG-based pairwise comparison method that outperformed FinBERT in portfolio backtesting.

B.Sc. in Physics (minor: Computer Science) · Bonn, Germany

• Bachelor's thesis (grade 1.3): Quantum Ising models on a Quantum Computer — designed and optimized gate sequences in multidimensional models with odd boundary conditions.