5 posts
Machine-reading the Shanhaijing's 26 mountain chains found 63 contradictions: 28 confirmed, 17 suspected, and one 429% gap that was ours, not the book's.
The Shanhaijing build turns 32,316 characters into 2,141 entities and 7,113 relations — 88.3% auto-extracted by regex rules, zero by LLM (project data, 2026).
A 429% distance gap in the Shanhaijing looked like a dropped 萬. It was the parser: the source reads 12,350 li, the pipeline logged 2,350 (project KG, 2026).
We trained LightGBM on KKBOX's Kaggle dataset and achieved a 0.792 average AUC predicting repeat listens. K-prototypes clustering grouped users and songs by similarity.
Average job postings draw 250 resumes (iCIMS, 2024). We processed 1.06M resumes on Hadoop with logistic regression to rank the top-100 candidates per role.