fortunehove5-sys/taura-ai

Voice-first multilingual AI decision-support infrastructure for inclusive agricultural, climate and financial information access in Zimbabwe. # Taura AI **A voice-first AI advisory agent for financial inclusion and climate resilience in Zimbabwe's informal economy.** *"Taura" — Shona for "speak."* Submitted for the **2026 AI For Impact (AI4I) Challenge — Track 3: Development**. See the full written proposal: `Taura_AI_Proposal.pdf` (in the parent submission package) for problem definition, strategic alignment, roadmap, compliance/risk analysis, and sustainability plan. This repository is the working reference prototype for that proposal. It demonstrates the product logic end-to-end — consent handling, multilingual (Shona / Ndebele / English) intent understanding, retrieval-grounded responses, and a human-escalation path — over three simulated channels (voice call, USSD, WhatsApp) plus a browser-based demo, all running fully offline against **synthetic sample data**. > **What this is / is not:** This is an architecture and product-logic > prototype, not a production deployment. Real ASR wiring for Meta MMS and a > fine-tuned Whisper checkpoint is implemented in `src/taura/asr.py` (see > below), but has not been run against real audio in this environment — no > network access to the Hugging Face Hub, no GPU. Text-to-speech and live > institutional data feeds remain stubbed integration points, clearly marked > — see `docs/ARCHITECTURE.md` for exactly what to > swap in and where. No real user data, and no real AMA/MSD/EcoCash data, is > used anywhere in this repository — see > `docs/DATASET_STATEMENT.md`. --- ## Quickstart ```bash python3 -m venv .venv source .venv/bin/activate # Windows: .venv\Scripts\activate pip install -r requirements.txt pytest -q # 57 tests, all offline, ~0.4s ``` ### Run the web demo (recommended for a first look) ```bash uvicorn backend.app:app --reload --port 8000 ``` Then open **http://localhost:8000** — a WhatsApp-style chat window on the left, and a "grounding inspector" panel on the right showing the detected language, classified intent, e …