Voice-first.
ASR and TTS are core, not bolt-ons. End-to-end speech to artifact.
Voice-first agentic instruct layer on TamilLM. Built for the uncle who speaks but never types.
The target user does not type.
The uncle in Kanyakumari, the farmer in Thanjavur, the grandmother in a Tirunelveli village - they speak. Text-first interfaces have failed these users for two decades.
Mugil is the voice-first agentic layer that sits on TamilLM and turns the substrate into something that actually shows up when a Tamil speaker needs something done.
"இப்போ எனக்கு ஒரு invitation வேணும்ன்னு print பண்ண - kovil thiruvizha-ku."
Speaking Kanyakumari Tamil with natural code-mix - he doesn't type, doesn't switch to English, doesn't adopt literary Tamil. If Mugil can hear his dialect, ask the right follow-ups in his register, know the cultural conventions without being told, and hand back a printable file by voice, end to end, then we've built something that doesn't exist today.
ASR and TTS are core, not bolt-ons. End-to-end speech to artifact.
Tirunelveli and Kanyakumari Tamil first-class. Not Chennai-Madurai default.
Chain-of-thought stays in Tamil. No English-thinking model in Tamil clothing.
Invocation order, kinship, panchangam - known, not retrieved.
No sorting by caste, religion, gender. Just does the job.