Engineering
Beyond apps, I build and operate a few systems that run in the background. This page describes what they do and what they are built from.
The source-citing question answering work runs on research hardware at Yuanpei University of Medical Technology, where I am a student.
Language models answer fluently even when they do not know. The fix is architectural, not model-level: an answer must be assembled from passages that actually exist, and each passage carries its address.
The reader clicks and sees the original. Citations are not a small footnote at the end — without them the answer is not finished.
documents ingested
Counted directly in Postgres on the ypurag machine · 2026-09-11
passages chunked and embedded
Counted directly in Postgres on the ypurag machine · 2026-09-11
questions logged
Question log on ypurag, 12 Aug – 10 Sep 2026 · 2026-09-11
Two translation bots running on LINE, serving real users every day. They translate between Vietnamese, English and Traditional Chinese.
followers on the main account
LINE Official Account statistics for the MonStudio OA, recorded in the operations log · 2026-09-08
LINE calls to the bot in one measured day
MonStudio bot on Appwrite Cloud, LINE webhook calls in 24 hours · server health report · 2026-09-10
median run time per call
Same bot, median run time across all calls in those 24 hours · server health report · 2026-09-10
A four-way translation API between Vietnamese, English and Taiwan-style Traditional Chinese, running on a model hosted in the room.
The model was chosen by measurement, not by name: the one with Taiwan in its name scored worst of the three tested.
score on a 0–100 scale, four directions weighted equally
Internal evaluation: 52 samples × 3 models, temperature 0, test set frozen before the run · 2026-09-08
median response time
Same evaluation, median over 52 samples · 2026-09-08
A smoke-detection model running on a live video stream, tuned to run on small hardware rather than needing a large server.
The camera-ingest server is built but has not been connected to any building's real cameras, so its event log is still empty. This project is unfinished.
processing time per frame
Single stream, preprocessing included, measured on an RTX 5070 Ti · 2026-09-11
share of smoke images caught
Test set of 707 real camera images: 153 with smoke, 554 easily confused · 2026-09-11
false-positive rate
Same 707-image test set · 2026-09-11
model size, nano variant
Nano variant after layer fusion; 11 MB ONNX file · 2026-09-11
If you have a problem close to any of the above, send me a message.
How to reach me