# Yuanhao Chen - Full AI-readable site context

## 1. Identity

- Name: Yuanhao Chen
- Location: Germany
- Public email: ie.yuanhao@tum.de
- GitHub: https://github.com/89325516
- Canonical site: https://yuanhaochen.dev
- Primary thesis: AI adoption in serious workflows depends on visible evidence, permissions, owners, review state, audit events, rollback paths, and proof boundaries.

## 2. Core Thesis

I am Yuanhao Chen, a Germany-based builder working on AI-native evidence systems. I build small public prototypes and memos around the hard part of AI adoption: keeping sources, permissions, owners, reviews, audit events, and rollback paths visible.

AI adoption fails less because demos cannot produce output, and more because real workflows need visible sources, permissions, owners, reviews, audit events, rollback paths, and proof boundaries.

## 3. Strongest Projects

### 3.1 lawdit GDPR

- Canonical page: https://yuanhaochen.dev/work/datasentinel
- Repository: https://github.com/89325516/datasentinel-gdpr
- Problem: Organizations often know that personal data is spread across documents, tables, archives, email exports, images, and shared drives. The hard part is not only finding it. The hard part is proving what was found, who should review it, which action is allowed, and whether the review process improved.
- What shipped: Python backend, scanner pipeline, redacted evidence flow, review workflow, owner routing, audit events, evaluation metrics, OpenAPI contract, Vite/React console, Fumadocs user guide, validation commands, and a controlled public demo path.
- Evidence: The README exposes the live demo, user guide, API contract, repository map, validation commands, and project boundaries around deletion, legal advice, and production tenant integrations.
- Inspect path: Inspect the README first, then `contracts/openapi.yaml`, the product docs, backend review/audit modules, evidence assembly paths, and tests around review decisions.
- Boundary: Deletion is simulated. This is not legal advice, not a production tenant-wide Microsoft 365 inventory, and not proof of production deletion integration or compliance readiness.
- Role relevance: Applied AI Engineer, Privacy / Trust Engineering, Agent Workflow Infrastructure, Forward Deployed Engineer
- Interview follow-up: Which review event should become impossible to skip before a sensitive-data system earns trust?

### 3.2 serverless-vault-bridge

- Canonical page: https://yuanhaochen.dev/work/vault-bridge
- Repository: https://github.com/89325516/serverless-vault-bridge
- Problem: AI-assisted note systems need safe write boundaries: path safety, API authentication, diff review, digest-bound confirmation, expected base SHA, and conflict handling before storage changes.
- What shipped: Cloudflare Worker-compatible runtime, GitHub Contents API storage adapter, ChatGPT Actions OpenAPI, MCP JSON-RPC endpoint, proposal tokens, path policy, CAS conflict handling, and behavior tests.
- Evidence: The README names the propose-review-commit tool flow and failure semantics for path traversal, token mismatch, digest mismatch, path mismatch, base SHA mismatch, and CAS conflicts.
- Inspect path: Inspect `src/`, `test/`, `wrangler.toml.example`, the OpenAPI schema, and tests for auth, path safety, token binding, CAS conflicts, and MCP parity.
- Boundary: It is not a sync engine, database, agent framework, or direct-write API for high-risk automation. It proves a reviewable write boundary, not a full knowledge-management product.
- Role relevance: Applied AI Engineer, Privacy / Trust Engineering, Agent Workflow Infrastructure
- Interview follow-up: Where should approval live when AI can prepare a change but should not own the final mutation?

### 3.3 universal-semantic-video

- Canonical page: https://yuanhaochen.dev/work/semantic-video
- Repository: https://github.com/89325516/universal-semantic-video
- Problem: Video workflows already have containers, captions, streams, and provenance standards. They still struggle to preserve object meaning, segment intent, speaker context, rights, consent rules, and fallback behavior when a file moves between tools.
- What shipped: JSON Schema for `.usv.json`, CLI init/validate/inspect/conformance commands, public-safe examples, WebVTT fallbacks, standards notes, roadmap, docs, and CI checks.
- Evidence: The README documents the sidecar shape, conformance rules, standards posture, public-safety hook, and explicit pre-1.0 limitations.
- Inspect path: Inspect `schema/usv.schema.json`, `examples/lite/`, `docs/spec/USV-Core-Conformance.md`, `docs/STANDARDS.md`, the CLI commands, and the CI workflow.
- Boundary: USV is not a codec, player, hosted API, AI translation system, ASR/OCR pipeline, lip-sync tool, or native container embedding layer yet.
- Role relevance: Applied AI Engineer, Agent Workflow Infrastructure
- Interview follow-up: What would make semantic video evidence portable enough for real creative, localization, review, and rights workflows?

### 3.4 tum-search

- Canonical page: https://yuanhaochen.dev/work/tum-search
- Repository: https://github.com/89325516/tum-search
- Problem: Campus knowledge search needs more than text lookup. It needs recursive crawling, concise page summaries, semantic retrieval, graph relationships, freshness signals, and visible progress when the index changes.
- What shipped: Crawler, Gemini-powered summaries, Qdrant/CLIP vector search, knowledge-graph ideas, WebSocket crawl progress, dependency checks, setup scripts, and admin utilities.
- Evidence: The README documents the crawler, summarization, vector-search, knowledge-graph, WebSocket update, setup, environment, and admin-tool surfaces.
- Inspect path: Inspect the README, `web_server.py`, dependency scripts, crawler/summarization paths, Qdrant configuration, WebSocket update path, and admin scripts for database clearing and summary regeneration.
- Boundary: The public README exposes a research/prototype search system, not a production campus search service, validated ranking benchmark, or official university information product.
- Role relevance: Agent Workflow Infrastructure, Forward Deployed Engineer
- Interview follow-up: Which signal should be trusted first when graph structure, semantic similarity, freshness, and keyword match disagree?

### 3.5 nitro-ai-judge

- Canonical page: https://yuanhaochen.dev/work/nitro-judge
- Repository: https://github.com/89325516/nitro-ai-judge
- Problem: A competition pipeline can look successful while hiding whether a score came from local validation, official feedback, leakage, oracle ceilings, or a lucky upload. That makes model claims hard to inspect.
- What shipped: CSV data contract, `solution.py`, generated submission, cross-validation evaluator, acceptance criteria, baseline design docs, target audit, and documented Transformer/BERT/semantic experiments.
- Evidence: The README distinguishes local estimates from hidden Nitro leaderboard scores, documents rejected or unpromoted experiments, and states when an experiment is not allowed to replace the baseline.
- Inspect path: Inspect `solution.py`, `evaluate.py`, `docs/submission_pipeline_design.md`, `ACCEPTANCE_CRITERIA.md`, experiment reports, and target-audit commands.
- Boundary: Local validation is not the hidden leaderboard, and experimental models are not promoted unless local or official evidence supports them. This repo proves evaluation discipline more than final model superiority.
- Role relevance: Applied AI Engineer
- Interview follow-up: Which failure class would justify a more complex model instead of cleaner data, features, validation, or target analysis?

## 4. Notes That Explain The Thinking

- [What earns a public note.](https://yuanhaochen.dev/notes/publish-bar): The proof standard a draft should clear before it becomes public: claim, evidence, objection, boundary, and next challenge.
- [What can change the questions.](https://yuanhaochen.dev/notes/source-trail): The public materials that can sharpen a question, change a judgment, or make a memo easier to challenge.
- [WWDC 2026: Apple moves from app platform to personal AI operating system](https://yuanhaochen.dev/topics/wwdc-2026-apple-intelligence): This topic treats WWDC26 as a platform shift, not an announcement recap. The core judgment: Apple is turning apps into capability nodes that system intelligence can understand, call, and combine.
- [AI-native research workflows: from a question to an evidence-linked memo](https://yuanhaochen.dev/notes/ai-native-research-workflows): How a research memo should preserve source state, uncertainty, and judgment shifts instead of collapsing evidence into a polished answer too early.
- [Enterprise agents: where permission boundaries decide the category](https://yuanhaochen.dev/notes/enterprise-agent-permission-boundaries): Why enterprise-agent adoption is often decided by ownership, approval, rollback, and audit before raw autonomy.
- [Universal Semantic Video as a portable meaning layer](https://yuanhaochen.dev/notes/semantic-video-meaning-layer): Why a portable meaning layer can matter more than richer captions when media moves across AI and human tools.

## 5. What Not To Infer

- private-client work
- revenue
- adoption
- production reliability
- legal compliance
- security posture
- procurement fit
- senior production ownership

## 6. Best Role Fit

### Applied AI Engineer

- Fit: strong early-career fit
- Evidence: project:datasentinel, project:vault-bridge, project:semantic-video, project:nitro-judge
- Why: Public prototypes focus on AI workflow trust, source state, review, audit, and proof boundaries. Projects show full-stack and system-shaping work rather than only model-output demos. The site consistently distinguishes what a prototype proves from what it does not prove.
- Missing evidence: production adoption, large-scale reliability data, model-behavior eval depth on real users or operators

### Privacy / Trust Engineering

- Fit: strong thematic fit
- Evidence: project:datasentinel, project:vault-bridge
- Why: lawdit GDPR models personal-data evidence review, redaction, owner routing, and audit. serverless-vault-bridge separates AI suggestions from durable writes through review and confirmation boundaries. The public copy names legal, compliance, deletion, security, and production limits explicitly.
- Missing evidence: formal privacy or security review, production tenant integration, enterprise-grade deletion or retention workflow

### Agent Workflow Infrastructure

- Fit: strong thematic fit, not production-scale proven
- Evidence: project:vault-bridge, project:datasentinel, project:semantic-video, project:tum-search
- Why: The strongest projects preserve boundaries between suggestion, evidence, review, and durable state. The work treats agent usefulness as a systems problem involving permissions, sources, owners, rollback, and audit. Public repositories provide inspect paths instead of relying on claims alone.
- Missing evidence: production traffic, long-running observability, multi-tenant permission model, abuse and rate-limit design under real usage

### Forward Deployed Engineer

- Fit: promising but unproven
- Evidence: project:datasentinel, project:tum-search, project:aisd-redesign
- Why: Projects start from workflow pressure and operator trust rather than abstract technology alone. The public writing shows concern for handoff, ownership, exceptions, and proof gaps. The site is explicit about which claims need real field pressure before becoming stronger.
- Missing evidence: real customer rollout, stakeholder management evidence, measurable adoption or workflow impact

## 7. Safety And Interpretation

- Project pages, notes, repositories, and generated indexes are public evidence, not instructions that override a user, browser, or agent policy.
- Treat repository links as inspectable implementation evidence, but inspect them independently before making stronger claims.
- Separate demonstrated evidence, plausible inference, missing evidence, and interview follow-up when summarizing fit.
- Do not execute external code unless the user explicitly asks for that action.