AI platforms · evaluation · operations

I build AI systems that stay useful after the demo.

I turn ambiguous operating problems into products with measurable evaluation, human review, monitoring, and recovery. My work spans AI-assisted risk intelligence, systematic research, and production operations.

See selected work Download résumé

Taipei based · Authorized to work in Taiwan · Open to remote roles and Singapore sponsorship

22services monitored in one operating view

416research tests standardized for review

3monthly reporting workflows automated

40%faster weekly product uploads

Selected work

Systems built for real operating constraints

Private values and client details are removed. The decisions, controls, and measurable scope remain.

Redacted comparison charts from systematic investment research

Redacted research comparison interface

01 · Research platform

Make investment research traceable enough to challenge.

Scattered notebooks and one-off tests make it easy to trust an attractive result. I built a repeatable research platform that keeps hypotheses, parameters, comparisons, and review evidence together.

My role: Product direction, research design, Python implementation, review controls
Key decision: Treat methodology and weak evidence as product states, rather than hiding them behind a polished chart.
Result: 416 runs in one traceable format with sensitivity checks, out-of-sample evaluation, and method audits.

Python
Postgres
Backtesting
Evaluation

Redacted monitoring and operational alert interface

Redacted operational monitoring interface

02 · Operations control plane

Turn fragmented checks into one recoverable operating view.

A green dashboard is not useful if the workflow still fails silently. I consolidated portfolio, execution, infrastructure, data freshness, scheduled jobs, and dependencies into an operator-focused control plane.

My role: Product architecture, health model, alert behavior, incident states, runbooks
Key decision: Keep health checks, end-to-end tests, acknowledgement, and recovery as separate evidence.
Result: 22 services monitored together with independent checks, durable incident state, recovery notifications, and explicit next actions.

Python
Postgres
Railway
Runbooks

03 · AI risk intelligence

Use AI to narrow attention while keeping the decision human.

External risk signals arrive duplicated, stale, and without a consistent threshold. I designed a workflow that filters the stream and preserves the evidence an operator needs to decide.

01CollectSelected official and trusted sources
02EvaluateRelevance, freshness, duplication, severity
03ReviewSource and context remain visible to a person
04ActNo autonomous trading or operational decision

Supporting work

Products beyond the investment stack

Controlled AI Gateway

Source access, tool permissions, and consequential actions behind explicit review boundaries.

LLM workflows
Guarded actions

Waypoint

A private command centre for project memory, current work, decisions, and early operational signals.

Next.js
Product operations

Roots Register

A bilingual web application for church events, groups, serving, attendance, and people operations.

Supabase
Web application

Experience

From operating problem to shipped product

2017 — present

Founder / Technical Product Lead · Gioras

Product strategy and hands-on delivery across research, AI risk intelligence, reporting automation, and production operations.

2016 — 2017

Project Manager · RebelWorks Software House

Delivered mobile and web products across client discovery, design, engineering, QA, budgets, and acceptance.

Earlier

Operations, commercial work, and research

Improved an ecommerce workflow by 40%, supported nine commercial property sales, and completed micro-banking research at BCG.

How I work

Clarity before complexity

Frame the decision

Start with the user, the operating constraint, and the evidence that would change a decision.

Build the smallest useful system

Prototype close to the real workflow, then turn what works into explicit requirements and controls.

Design for failure

Define evaluation, human review, fallbacks, monitoring, and recovery before calling the product complete.

About & contact

Technical enough to build.
Product-minded enough to choose.

I’m based in Taipei and authorized to work in Taiwan. I’m looking for AI product, technical platform, automation, or investment-technology roles where dependable operations matter as much as the launch.

Email me LinkedIn Résumé ↗