I am Haipeng, an independent builder learning in public. I make small AI tools that people can read, test, and verify before trusting them.

My current focus is evidence-aware AI tooling: visual knowledge navigation, scientific-compute preflight, declared experiment readiness, Skill portability, and evaluation.

Evidence over unsupported claims. Useful releases over large promises.

An interactive, non-generative knowledge navigator for learning how declared bio/formulation relationships fit together.

• selecting a synthetic functional role reorganizes a particle constellation;
• explorable candidates move inward, while soft conflicts, missing conditions, and unknown evidence require explicit review;
• demo hard conflicts move to a non-selectable perimeter with the rule, condition, evidence, version, and review date exposed;
• effect hypotheses stay hidden while any selected relationship is unresolved;
• the bilingual, keyboard-accessible site runs without uploads, telemetry, persistence, user formulas, or a backend.

Open the live constellation

Two deterministic, local-first Agent Skills for checking research plans before any separate execution or experiment:

• plan-research-compute reviews sanitized metadata, governed data locations, an abstract DAG, pinned tools, and bounded resources without reading data or launching cloud jobs;
• check-formulation-readiness checks user-supplied arithmetic, units, bounds, process order, evidence scope, and experimental gates without generating or optimizing a formula;
• default formulation reports are ID-only, and a ready result never means feasible, safe, stable, effective, or compliant;
• ships as the audited v0.1.0 release.

A read-only static auditor for Agent Skills. It catches broken resources, escaping paths, undeclared requirements, and Codex/GitHub Copilot portability risks before installation or publication.

• never executes scripts from the Skill being audited;
• emits deterministic text, JSON, or SARIF;
• includes adversarial tests, public-scope gates, CodeQL, and release audits;
• ships as a pinned, installable v0.1.0 release.

• inspectable Agent Skills with explicit inputs, outputs, and limits;
• visual knowledge tools that keep unknown and conditional relationships explicit instead of inventing a confident answer;
• local research preflights that keep private data and real formulas out of public repositories;
• deterministic checks for failures that prompts alone cannot catch;
• guardrails for agent cost, fan-out, concurrency, and run budgets;
• practical documentation that separates tested evidence from assumptions.

• Preview before install.
• Audit before explanation.
• Keep target access read-only by default.
• Publish synthetic tests, exact limitations, and reproducible checks.
• Improve from small counterexamples supplied by real users.

1. Improve the released projects with community-submitted synthetic failures and reviewed knowledge-model edge cases.
2. Expand BioFormulation Constellation's curation tooling before accepting any real-world assertion.
3. Build an agent-run-budget-guard Skill for preflight limits on calls, tokens, concurrency, and elapsed time.
4. Explore focused pet-AI and research-operations Skills without publishing private product or formulation work.

If this direction is useful, star or watch BioFormulation Constellation, AI Research Preflight, or Professeur AI Skills, then open an issue with the smallest synthetic failure you can share.

我是海鹏，也使用 Professeur Haipeng 这个名字。我正在公开学习和构建可检查、可测试、可复用的 AI 工具。当前公开项目包括生物配方知识星图、科研计算预检、配方实验就绪检查和 Skill 静态审计；真实配方、私人研发资料和原始科研数据不会进入公共仓库。