I want to walk away from this engagement with a fully functioning, phone-based Hindi conversational companion that young adults will actually enjoy calling. The scope is an MVP that we can build, test and iterate on within 4–6 weeks, so I’m deliberately steering us toward proven Speech-to-Text, Text-to-Speech and LLM APIs rather than reinventing the wheel. What matters most is how naturally the system sounds in Hindi and how smoothly it moves through the full spectrum of emotions—light banter when the caller is relaxed, calm reassurance when they are stressed, and everything in between. The assistant must handle interruptions, follow-ups and context switches without awkward delays or robotic phrasing. You’ll own the end-to-end flow: • Telephony layer (Twilio, Plivo or similar) that answers a live call, streams audio and hands it to your STT pipeline. • Real-time STT in Hindi (Google, Azure, Whisper v3, etc.) with low latency. • Prompt engineering & memory for the chosen LLM so it maintains context over a multi-turn conversation. • High-quality Hindi TTS (Amazon Polly, Azure Neural, Google Wavenet, etc.) that returns voice with proper intonation. • Simple admin dashboard or logging endpoint so I can review transcripts, latency metrics and error traces. Acceptance criteria 1. A test phone number I can dial from my mobile in India and hold a five-minute conversation that stays coherent, context-aware and emotionally appropriate. 2. Average round-trip latency (caller finishes speaking → bot replies) under 2.5 seconds. 3. All code, prompts and infra scripts delivered in a Git repo with a one-command deploy to my AWS or GCP account. Tech stack is flexible—Node.js, Python (FastAPI), or Go are fine as long as the build is clean and documented. If you’ve shipped real-time voice bots, especially in Hindi, let’s make this happen.