https://www.hridaysaha.com/post/rag-chunking-the-bug-that-lived-inside-a-page-footer 2026-07-16 https://static.wixstatic.com/media/78a230_e41df0d1d3504b2b93a3d079bdb531a3~mv2.png A policy page showing lettered clauses that each nest a roman numeral "i." sub-item, and a repeating registration footer highlighted at the bottom, the two parsing traps in the document. https://www.hridaysaha.com/post/rag-evaluation-the-fix-i-was-proud-of-and-the-measurement-that-killed-it 2026-07-16 https://static.wixstatic.com/media/78a230_1308e9374c40400f88fca94d56978d1d~mv2.png A two-stage design that felt obviously better, measured across 34 questions and losing: citations verified fell from 100 percent to 61.8 percent, and hallucinated citations rose from 0 to 13 of 34. https://www.hridaysaha.com/post/roc-vs-precision-recall-curves-when-accuracy-lies-to-you 2026-07-12 https://static.wixstatic.com/media/78a230_494b69c16b3a4a8f93505912ec3dbebc~mv2.png https://www.hridaysaha.com/post/cve-to-cwe-mapping-retrieval-and-rank-pipeline 2026-07-12 https://static.wixstatic.com/media/78a230_fdea56070c7943b3b654b35384634250~mv2.png Pipeline diagram showing CVE description flowing through BM25/TF-IDF retrieval, Sentence-BERT reranking, and LightGBM ranking to produce ranked CWE suggestions https://www.hridaysaha.com/post/why-my-model-couldn-t-predict-growth 2026-07-28 https://static.wixstatic.com/media/78a230_6d0da93e25bd4c32b020a415e1a4ca5c~mv2.png Why my model couldn't predict growth: tree models can't extrapolate, so forecast a bounded one-step change and reconstruct the reported level — the FinSight approach https://www.hridaysaha.com/post/point-in-time-data-leakage-when-historical-data-isn-t-historical 2026-07-27 https://static.wixstatic.com/media/78a230_693cc50e09a14c71990a8daf489fbec8~mv2.png Point-in-time data leakage, explained: reconstruct every value as it was known at filing time so the future never leaks backward — the FinSight pipeline https://www.hridaysaha.com/post/when-baseline-beats-machine-learning 2026-07-28 https://static.wixstatic.com/media/78a230_22029f92271f47349896dade065bfd5e~mv2.png When the baseline won: on free cash flow, a one-line seasonal baseline beats machine learning, so FinSight serves the baseline and labels it as such