Sina — Bilingual Clinical Intake Voice Agent

Patients don't speak in one language on demand. Across Oman and the wider Gulf, an intake conversation moves between Arabic and English mid-sentence — a symptom in English, a history in Arabic, sometimes both in the same breath. Most voice tools force a choice. Sina doesn't. Sina is a real-time bilingual voice agent for clinical intake. A patient speaks in Arabic, English, or a mix, and Sina transcribes with AssemblyAI's streaming API in Medical Mode, extracts a structured summary (symptoms, medications, allergies, duration, severity), replies in a native Omani Arabic voice or an English voice, and escalates to a human interpreter when needed. The model doesn't get the final word on safety. A reviewed list of red-flag symptoms — chest pain, breathing difficulty, stroke signs, severe bleeding, seizure — is matched against the raw transcript in plain Python, not the model's paraphrase. If the LLM misjudges a chest-pain case as low-urgency, this layer overrides it anyway; it can only add an escalation, never remove one. Escalation doesn't cut the conversation short. It fires as a one-time banner while the conversation keeps going and asks its normal follow-ups, handing off only once intake is complete. The bilingual experience goes beyond the AI's replies. A first-load popup lets the patient choose Arabic or English, translating the whole interface with right-to-left rendering. Languages can switch mid-conversation without losing anything gathered, and low-confidence transcriptions get a real yes/no instead of a silent guess. Real-voice testing drove the hardest decisions: code-switching garbled transcription on ambiguous audio, so Sina pins one language per session; and ElevenLabs had no free Arabic voice, so Arabic speech runs on Azure's native Omani voice instead. Sina doesn't diagnose or replace emergency care — it exists so clinicians get accurate information regardless of language, and nothing urgent slips through on false AI confidence.