Final-year Mechanical Engineering student at National Cheng Kung University, moving into research on trustworthy AI agents for finance — making AI-generated strategies and analysis reproducible, verifiable, and auditable.

The question I keep coming back to: when an AI can generate code, generate analysis, and even operate a system directly, how do we confirm its output is trustworthy? In finance this bites hard — a backtest that overstates performance, or an analysis that quietly invents a number, costs real money.

I work agent-first: I design the systems, specify the validation that decides accept-or-reject, and judge the results; AI coding agents (Claude Code / Codex) do the implementation. The part I care about — and want to research — is the checking layer.

flowchart LR
    A["AI agent<br/>strategy · analysis · actions"] --> V{"Verification layer<br/>gates · provenance · holdout"}
    V -->|passes| OK["Trustworthy output"]
    V -->|can't prove| STOP["Reject / abstain"]

Trustworthy, auditable AI agents in quantitative finance, along two lines I'm prototyping. Each line has a working prototype and a measured result — including the ones that went against me.

• Strategy trustworthiness — can overfitting be caught reliably? Validation gates (trial registration · CSCV/PBO · DSR · single-use holdout · no-look-ahead contract) against a naïve-backtest baseline. On record: three pre-registered FAILs, adjudicated and published rather than buried.
• Analysis faithfulness — retrieval agents with a provenance verifier: every figure and citation must trace to a source, or the claim is blocked and the agent abstains. Measured, not asserted: 0% false positives on hand-built red-team suites, 0.06% on 10,000 pairs of a public benchmark nobody here wrote.

• trialgate — the validation gate as a zero-dependency, fully typed package on PyPI: trial registry · CSCV/PBO · Deflated Sharpe · single-use holdout lock. The working prototype of this research line. It has already rejected one of my own strategies.
• tw-stock-trading — that rejection, in full. Three pre-registered timing experiments on 21.2 years of Taiwan 0050 total-return data, net of statutory costs. Drawdown protection passed (26.3% vs 55.8%); the CAGR cost failed the pre-registered claim (4.91%/yr vs 10.21% buy-and-hold). Verdict: three FAILs, and per the stop rule the production runtime was never built. The honest answer was buy-and-hold.
• crypto-quant-signal — the same six gates on a running crypto testbed: spot, long-only, public data, daily signals, a 1000 USDT paper scoreboard. Holdout still sealed; a ≥3-month paper qualification runs until October before the pass/fail report. No API keys, no live orders — by product definition. A two-minute offline demo replays 713 decision cycles with no network.

• report-workflow — a deterministic gate layer between an LLM and the document it hands you: a number, quote or reference that is not in your sources never reaches the page, and you are told which sentence and why. The gate itself calls no model. Evidence, not adjectives: 88.6% recall / 0% false positives on a 44-case red-team suite (all 15 attack families caught at 100%; the surviving evasions are kept in the test set on purpose), and 0.06% false positives / 99.7% block precision out of domain on HaluEval's 10,000 QA pairs. On PyPI and shipped as an MCP server; 917 tests, CI green.
• legal-agent — a RAG pipeline where every citation is checked by code, and the checker is itself graded by seeding errors into otherwise-correct answers: 11,904 / 11,904 caught, 0 false positives. A time-sliced statute store cites the law in force at the event's date (as_of=), and the whole pipeline runs with no model, no key and no network — which is also how it is tested. My deepest system-design work, and the seed of this research direction. Live demo — about thirty seconds; it ships a pre-filled broken answer to check for yourself.

• otto — a local financial terminal an AI operates end-to-end over MCP, one claude mcp add away. Two things make it a testbed rather than a toy. Agent operability is a benchmark, not a tagline: a real headless agent runs plain-language tasks in hermetic sandboxes, graded programmatically on terminal state and artifacts, never by an LLM judge (claude-sonnet-5 20/20 on the 20-task suite, 2026-07-10 — the suite has since grown, so those figures are dated rather than current). And safety is structural rather than switched off: live trading and credential entry are unreachable through the surface the agent has, and refusal tasks are graded on state unchanged. It also keeps a scorecard of its own dated calls, and publishes it when the calls are bad.

• OpenRead — read the web and your PDFs in your own language without sending a word of it anywhere: Chrome's on-device translator by default (Ollama as a fallback engine), bilingual output under each paragraph, Traditional Chinese that reads like Taiwan. No account, no API key, no cloud.

Self-taught career-changer (mechanical engineering → AI / quant finance). Strong at decomposing problems, prototyping fast with AI tooling, and validation / risk thinking; seeking graduate research training to deepen the statistics and evaluation methodology behind all of it. 中文（母語）· English (C1) · 日本語（基礎）.

Since August 2026 I've been in a mentored remote internship in quantitative research and data engineering — the first work of this kind I'm doing to someone else's brief rather than my own.