Offline-first NLU & edge-deployable assistant models. Focused on intent classification, slot filling, and out-of-domain rejection for practical voice/text assistants. Building compact, state-aware transformer models that run locally without cloud dependency.