01 / built work
Things I built, tested, and kept inspectable — including the evidence and the parts that are still unfinished.
02 / lead evidence
7 published projects in the workspace.
Getting an AI to answer is easy. Knowing what it based the answer on is the interesting part.
I built a support-triage workflow that keeps the retrieved cases, drafted answer, grounding check, and human next step visible.
Open project
There is plenty of public data. How do we know where the number on screen came from?
A search and analytics project for Thai procurement data that keeps synthetic demos separate from a bounded, traceable snapshot.
Open project
Data can look clean at the start. Is it still the same data after it crosses five layers?
A local-first retail pipeline that moves synthetic data through Bronze, Silver, a blocking quality gate, and a warehouse before reporting.
Open project
If the model can give a label but is unsure, which reviews should go to a person?
A Thai review classifier that treats the label as a starting point and exposes cases that may need human follow-up.
Open project03 / more artifacts
Not every useful project needs to become a product.
A model can be right. How much can we honestly explain about why?
An educational study of model behavior on the Breast Cancer Wisconsin dataset, with bounded explanations and explicit limits.
Open project
If AI reads a receipt incorrectly, where should a person get the final say?
A local-first expense tracker that lets AI parse a receipt, but waits for a person to review and confirm before saving it.
Open project
How does one taxi file become a dashboard without letting bad rows quietly through?
A local-first trip-data platform that profiles, validates, models, and serves a small fixture while keeping rejected rows visible.
Open project