• Agent Runtime / Harness — bounded Agent Loop, versioned Tool Contracts, persist-before-execute, experiment governance, crash reconciliation, cancellation, and deterministic result validation.
• RAG Evaluation — frozen TechQA benchmark with 28,481 documents / 610 answerable queries; Dense Top-100 + rerank improved held-out Recall@5 from 64.4% → 72.5% (+8.1pp) and MRR@10 from 0.519 → 0.561.
• Backend Engineering — Python / FastAPI / PostgreSQL, async workflows, state machines, database migrations, automated tests, CI, and reproducible local environments.

A controllable, recoverable, and verifiable runtime for machine-learning experiments. It separates Agent decisions, human approval, physical execution, recovery, and deterministic result validation instead of letting an LLM directly control long-running training jobs.

Engineering evidence: fault-injection validation showed that a Coordinator restart could reconcile the original training Attempt without a second physical launch, with explicit cancellation paths for running and pre-launch work.

Python FastAPI PostgreSQL asyncio Agent Runtime Context Engineering Tool Calling SSE

An enterprise IT support backend combining document lifecycle management, RAG, controlled ticket creation, Human-in-the-loop approval, AgentOps tracing, and an offline evaluation pipeline.

Held-out DEV: Dense Top-100 + qwen3-rerank improved Recall@5 from 64.4% → 72.5% (+8.1pp) and MRR@10 from 0.519 → 0.561 on a frozen TechQA benchmark.

Python FastAPI SQLAlchemy ChromaDB RAG Rerank AgentOps Pytest

A Windows-first local desktop app for understanding computer usage, work-rest rhythm, reminders, and focus signals. It captures foreground-app activity, applies privacy processing and local classification rules, stores data in SQLite, and provides review workflows and desktop-companion surfaces.

Tauri Vue 3 TypeScript Rust SQLite Local-first

A leakage-controlled Agent Skill optimization and evaluation pipeline designed to distinguish reusable Skill improvements from public-task overfitting.

Evaluation design: frozen 60 / 17 / 10 Dev–Holdout–Final-Blind split, explicit no_skill vs with_skill paired runs, execution provenance, regression gates, and generalization checks.

Python Agent Skills BenchFlow SkillsBench Agent Evaluation Docker

• Codex History Restorer — local Windows tool for recovering Codex Desktop conversations that still exist on disk but no longer appear in the application.
• Bilibili Content — reusable CLI tool and AI Skill for subtitle extraction, validation, transcription, and structured downstream processing.

• Agent / LLM: Context Engineering · Tool / Function Calling · Human-in-the-loop · RAG · Retrieval / Rerank · Agent Evaluation
• Backend: Python · FastAPI · asyncio · SQLAlchemy · PostgreSQL · SQLite
• Engineering: Pytest · Alembic · GitHub Actions · Docker Compose · Git · Linux / Shell · Ruff · mypy
• Product / Desktop: TypeScript · Vue 3 · Rust · Tauri
• Research: PyTorch · causal discovery · causal representation learning · generative models · counterfactual generation

Master's candidate in Communication Engineering, researching causal structure learning for diffusion-based generative models, including causal discovery, causal representation learning, and counterfactual image generation.

First-author manuscript submitted to Neurocomputing.