Building reproducible AI systems from experiments to engineering deployment.

_{AI Undergraduate · 2027 Graduate — research → experiments → engineering → reproducible evaluation.}

Hierarchical reinforcement learning for quadruped robot navigation in MuJoCo — high-level PPO over a scripted controller, with BC, DAgger and curriculum experiments.

PPO Behavior Cloning DAgger Curriculum Learning MuJoCo

• Curve 20/20, multi-stage A→B 20/20, junction routing 40/40, 0 falls across these archived runs.
• Includes ablation studies, failure→diagnosis analysis and a reproducible evaluation protocol.
• 196 passed / 1 skipped, Pyright 0 errors (documented local run).
• Simulation-only (MuJoCo); branch selection and part of docking are scripted, final PPO weights require retraining.

Experiment-infrastructure project: chronologically replay RL training runs, flag training risk and validate experiment-data quality before trusting a result.

Chronological Replay Risk Analysis Data Quality Experiment Infra

• Configurable R0–R3 risk decisions (continue / watch / stop / tune / resize).
• 6 data-quality categories / 20+ checks; cached-vs-direct pipeline equivalence tests.
• 265 tests passed (documented local run); test gate runs on Linux CI.
• Analyzes robot-training logs — not market trading. Look-ahead isolation is unfinished and type-check debt is tracked openly in Limitations.

Risk-intelligence infrastructure for AI-native businesses and agents: Evidence → Risk → Action, with a deterministic risk engine behind a versioned API.

Deterministic Risk Engine TAI / CCI Versioned API Docker AI Agents

• Versioned deterministic TAI (5 components) and CCI (5 dimensions) plus confirmed-event integrity veto.
• Bearer-auth, bounded, idempotent versioned API; containerized with Docker.
• 67/67 tests passed (documented local run); External Alpha.
• LLMs assist evidence extraction and explanation only — deterministic rules own every authoritative score.

An open-source Agent Skill for structured product definition, feature tradeoffs and technical decision review, inspired by public Apple product practices.

Agent Skills Product Decision Systems Technical Review AI Tooling

• Turns product ideas, PRDs and technical proposals into explicit keep / merge / demote / remove / verify decisions, with concrete alternatives instead of general "make it simpler" advice.
• Encodes a reusable workflow: one-line product definition → simplification → end-to-end experience gaps → prototype validation.
• Four documented modes — product review, product definition, methodology Q&A, technical proposal review — running on one model-agnostic layout: SKILL.md, five references/, a review checklist and output templates.
• Separates public product history from methodology inference, and labels assumptions and unverified claims instead of asserting them.

Install — npx skills add Anhao1314/Apple-skill

_{Independent project — not affiliated with or endorsed by Apple Inc. Validated through structure, reference and packaging checks; no client installation or Agent behavior test has been run yet.}

• Run the full experiment loop — define baselines, ablations and controls, then read failure modes instead of only reporting the winning run (Go2W: BC-vs-init, DAgger and goal-conditioning ablations with explicit interpretation boundaries).
• Build reproducible evaluation pipelines — fixed protocols, archived reports and per-project validation records that state environment, scope and what was not proven.
• Debug data and training-quality problems — reward-drawdown / stagnation / restart risk items and six-category data-quality screening (RL Training Risk Replay).
• Lock behavior with tests — 196 passed / 1 skipped (Go2W), 265 passed (Risk Replay), 67/67 (FlowCredit), with honest accounting of skipped and advisory checks.
• Take a prototype to a testable system — versioned, schema-validated API, deterministic core, Docker packaging and documentation (FlowCredit), while keeping the LLM/rule boundary explicit.
• Package a decision method as a reusable Agent Skill — encode product and technical review into a repeatable workflow with explicit keep / merge / demote / remove decisions, evidence labels and reusable output templates (Apple Product Design Skill).

Python PyTorch MuJoCo Gymnasium Stable-Baselines3 NumPy Pandas SciPy scikit-learn Linux Git Docker

_{Also used in FlowCredit: JavaScript · Node.js · JSON Schema / Ajv.}

• Main thread across projects: Algorithm Engineering · Reinforcement Learning · Robotics · Decision Systems · Experiment Infrastructure · Reproducible Engineering · Agent Tooling.
• Seeking 2027 internship / new-grad roles in algorithm engineering, reinforcement learning, robotics or embodied-AI decision systems.
• I keep claim boundaries visible: simulation vs. real robot, post-hoc vs. predictive, deterministic vs. LLM — no result is presented as stronger than its evidence.