Featured work
All work →
Making KYC Resumable Without Starting Over
Autosave started as a way to keep KYC progress from disappearing when a session was interrupted. Once KTP and Selfie could be saved independently, the harder question was whether users still needed a fixed order. I tested the risk of a flexible flow and used the evidence to align the team before shipping it to production.
Designing a KYB Fast Track for 15,000 Merchants
An AstraPay partnership brought a roughly 15K SRC merchant cohort into KYB while the same Operations team, about three reviewers, continued handling BAU. The CMO asked Ops and Product to smooth the onboarding path. Instead of rebuilding KYB, I proposed a bounded fast-track model that reused referral-code infrastructure, moved clear cases to one-page single-pass review, and routed exceptions to deeper review. We validated the path on a whitelisted SRC sample, trained reviewers, then rolled it out over roughly 3–5 months.
Turning QRIS Into a Top Up Channel
QRIS Top Up started from a business request to make top up feel more immediate without always opening a banking app or exchanging account details. After benchmarking competitors and more interoperable QRIS patterns, I used the QR behavior already familiar across Indonesia as a practical bridge for several use cases, from friends funding each other to event admins distributing balance. With engineering, I designed the real-time post-payment states and kept each closed-amount payment tied to one bill so status and reconciliation stayed coherent. By 2026, QRIS Top Up represented 9.3% of AstraPay cash-in and ranked among the top two channels.
Testing When a Pedagogical Agent Actually Helps Learning
I designed a 2×2 mixed-factorial study with 157 participants to test whether a visible pedagogical agent helps psychomotor and cognitive learning tasks in the same way. The agent improved psychomotor performance and self-efficacy, but did not improve cognitive performance and slightly reduced cognitive self-efficacy. I also validated AI-based gesture scoring against manual ratings (r=.842) and checked the key findings with robust models when statistical assumptions were not clean.
