
Customer Behavior Analysis with YOLOv26
An end-to-end retail computer vision system that detects pre-purchase shelf behavior with a custom-trained YOLO model — mAP@50 60.91% across 4 interaction classes on 12,578 frames.

An end-to-end retail computer vision system that detects pre-purchase shelf behavior with a custom-trained YOLO model — mAP@50 60.91% across 4 interaction classes on 12,578 frames.

1st Place at the CiMSO Hospitality ERP Hackathon 2025 — a manager-centred analytics platform with 25 Chart.js visualizations, 45 KPIs, five languages and an LLM-powered insight assistant.

An automated NLP pipeline that scrapes Thai investor discussions from Pantip.com, links them to SET-listed tickers, scores sentiment with XLM-RoBERTa and serves a live Streamlit dashboard.
No projects match these filters.
Try another category or tag, or start again from the full list.
1
Conference paper
RSUCON 2026 research on how digital multitasking, cognitive load and technology use relate to student GPA. Python multiple linear regression (n=196) found Technology Use the sole significant predictor (β=0.137, p=0.004).
0
Written by other researchers — my notes, not my work.
Papers I study, with my notes, will be listed here.
LÉOR, a wearable personal-safety concept — smart wristband and companion app — won 1st place; I shaped the product narrative, the Figma prototype and the live pitch.

How a manager-first analytics dashboard — 25 charts, 45 KPIs, five languages and an LLM assistant — took 1st place out of 11 finalist teams.
W422026
Week of
Concept
mAP@50 only asks whether a predicted box overlaps the truth by half; mAP@50–95 averages ten stricter overlap thresholds, so it rewards tight boxes. When localisation matters, report both.
Observation
A macro-averaged score can look fine while one rare class is close to zero. Per-class precision and recall show where the model actually struggles.
W412026
Week of
Concept
Correlation between sentiment and next-day returns only says the two move together after a delay. A Granger-style test or a hold-out period is needed before calling sentiment predictive.
Learned
Comparing each day with a rolling mean and standard deviation flags spikes without a fixed threshold that goes stale as volumes change.
Reading
Re-read how XLM-R was trained on text in 100 languages and why one shared vocabulary helps lower-resource languages such as Thai.
W402026
Week of
Tool
Cron schedules use UTC and may be delayed when runners are busy, so a pipeline should never assume it starts at an exact minute.
Learned
document.startViewTransition() snapshots the page, lets you swap the visible panel and animates between the two states with plain CSS.