YeasirArafat

Nine methods, including re-implementations of published baselines, on CIC-IDS2017, CSE-CIC-IDS2018 and UNSW-NB15 over five seeds. Mine matched the strongest baseline on macro-F1 on both CIC datasets while flagging fewer benign flows on every dataset.

On a 62-question gold set the system invented nothing — eight of those were stopped by a retrieval score floor before any LLM call was made. Answers that do land carry page-level citations.

363 backend and 35 Flutter tests. The audience filter runs inside the SQL before vector search, because post-filtering leaks a document's existence through the not-found path.

Calibrated to an expected calibration error of 0.020. I rejected a 0.94 AUC model after a leakage audit showed it was 0.18 AUC worse on people who had never been screened — exactly the target user.

143 backend, 55 widget and 28 screen goldens across a 19-screen Flutter client in two languages and two themes.

Rice, potato, tomato and corn, running INT8 TFLite with no network at inference. Reported against a held-out field test set rather than a random split of lab images.

5,843 NHANES rows with waist circumference beat 437,000 pooled rows without it — 66.9% against 52.9%. Sample quality beat sample volume.

Telegram, WhatsApp, Messenger and web chat normalised into a single payload, then routed by a hybrid keyword and LLM orchestrator emitting structured intent and confidence.