Wei-Ting Liu
NTU Information Management + Trans-disciplinary Bachelor Degree.
Full-stack engineer with an AI-native mindset, building data-driven products.
Open to every chance and place to learn.
Full-stack engineer moving into reinforcement learning R&D, currently designing a dual-loop RL system at ABConvert that trains agents to optimize Shopify stores while closing the sim-to-real gap. In parallel, running two computational-epidemiology studies: a multi-agent RL benchmark for decentralized dengue alerting across Brazil's 26 states, and a 10.2-billion-row measurement of age-mixing structure in Seoul's mobility data. Both released as citable replication archives.
Grounded in shipped product work: migrated a 10-page marketing site to Next.js at a 97 Lighthouse SEO score, built internal analytics dashboards that cut manual metric lookups by ~70%, and led AI adoption workshops that moved 80% of a research center's staff from chat-based AI use to building their own tools. B.B.A. student in Information Management and the Trans-disciplinary Program (College of Innovation) at NTU.
ABConvert — Reinforcement Learning R&D Intern
Shopify A/B Testing SaaS
- Designed a dual-loop RL system that trains agents to optimize Shopify stores, aiming to build an accurate, real-world-calibrated simulator that closes the sim-to-real gap.
NTU Insight Center — AI Intern
- Developed an internal service-design product and led AI adoption workshops, moving 80% of staff from chat-based AI use to building their own website products.
ABConvert — Software Engineer Intern
Shopify A/B Testing SaaS
- Migrated abconvert.io (10 pages) from Webflow to a self-hosted Next.js app, reaching a 97 Lighthouse SEO score and enabling custom interactive components.
- Built internal dashboard features (experiment research, visualization, analysis) from a cross-team needs survey, eliminating ~70% of manual metric lookups for Customer Success and the CEO.
ABConvert — Startup Generalist Summer Intern (Product/AI)
Shopify A/B Testing SaaS
- AI-native generalist: LLM-assisted workflow to compress spec → implementation loop; shipped Next.js + TypeScript frontend (SSR/ISR, streaming UI) at startup pace.
Multi-Agent RL for Decentralized Epidemic Surveillance
Advised by Wayne Lee · 10.5281/zenodo.21619163
- Built a CTDE multi-agent RL framework (MAPPO/QMIX/IPPO/CQL on PettingZoo) modeling Brazil's 26 states as agents alerting on dengue from 258 weeks of case counts and Google Trends, benchmarked against an information-parity threshold rule.
- Showed the agents lose systematically on historical replay and decomposed the gap into reward misalignment (partly recovered by a gated false-alarm penalty) and residual on-policy sample inefficiency; the same agents do learn control on action-coupled SIR/SIS environments, localizing the failure to task specification rather than to RL. Released a 107-checkpoint replication archive; manuscript in preparation.
Age-Mixing Structure in Aggregated Urban Mobility Data
Advised by Wayne Lee · 10.5281/zenodo.22152089
- Measured age-assortative mixing across 10.2 billion rows of Seoul's Living Migration release (79 months, 424 administrative dong, 16 age bands), pairing the near-zero reading with a synthetic-world calibration that fixes the instrument's detection limit.
- Found the co-arrival matrix within 0.00202 of proportionate mixing in every month — a factor of 33–35 below a national contact survey — while the pipeline recovers 92–101% of true assortativity on synthetic controls; verified by a mutation-tested citation audit, a determinism harness, and an independent re-implementation agreeing on 436/436 checks.
National Taiwan University
B.B.A. in Information Management · Double Major: Trans-disciplinary Program (College of Innovation) · Leadership Program
olcli — Overleaf CLI
Open-source contributor
TypeScript, Node.js. Contributor to olcli, a CLI that pulls, pushes, and compiles Overleaf projects from the terminal (175+ stars, distributed via npm and Homebrew). Contributed olcli diff, a content-level diff of the local tree against the live remote project — 900+ lines with unit and e2e tests, refactoring shared file scanning into a reusable module — and a fix for a zip-slip path-traversal vulnerability in project archive extraction.
Next.js 15, TypeScript, React, Tailwind CSS, Spotify Web API, shadcn/ui. Real-time multiplayer music guessing game with 70,000+ monthly active users; host pastes any Spotify playlist URL and the app fetches tracks via Spotify Client Credentials, running a round-by-round quiz with audio clips, live scoring, and answer validation.
ABConvert.io Website Migration
Next.js, TypeScript. Migrated abconvert.io (10 pages) from Webflow to a self-hosted Next.js app, reaching a 97 Lighthouse SEO score and enabling custom interactive components.
Open to internship and full-time opportunities in SWE, AI/ML, and product engineering.