Raihan
Author ProfileClaim This Author Profile
Prove ownership by publishing #HashtagPLUS and this profile link on your author page or an article under your byline. A moderator or admin will review the request before it merges into your real HashtagPLUS username.
🌐 dev.toSource
Frozen benchmark over nine yearly FAR revisions: dateless RAG answers from the wrong year 80% of the time; year-pinning fixes it but breaks change questions.
🌐 dev.toSource
RAG over nine yearly FAR revisions: plain hybrid leaks wrong-version answers 80% of the time (material 32%); version handling fixes it without hurting recall.
🌐 dev.toSource
Trained sparse autoencoders per layer on llm-jp-3-1.8B: JA-EN sharing rises then falls with depth (names, then concepts, then script); pre-registered causal tests come back weak-JA, null-EN.
🌐 dev.toSource
Built an all-local streaming JA-EN speech translator and measured the verb-final wait cost: EN-JA prefix translations survive ~13x worse than JA-EN at matched prefixes (same backend).
🌐 dev.toSource
Voice-first Japanese tutor agent that works end to end — but its correcting brain predicts particle on 285/300 rows and never the other types. A pre-registered, measured failure.
🌐 dev.toSource
From Dev Community: I built the first open benchmark for federal contracting AI. Here's what it shows about frontier LLMs.
🌐 dev.toSource
From Dev.to - python: Three small models for healthcare intake — and what shipping all three taught me
🌐 dev.toSource
From Dev.to - python: Where small models beat frontier LLMs (and where they don't): a 125M PHI detector
🌐 dev.toSource
From Dev RSS Feed: Matching frontier LLMs at 22 lower latency: a 184M-parameter intent classifier for healthcare text