CX AI - Conversational Data Assistant

CX AI — Conversational Data Assistant Project Requirements · Fixed-Bid (Upwork) We want an AI assistant that lets CattleXpert customers ask questions about their own feedlot data in plain language — typed or spoken — and get accurate, real-time answers. The user asks a question; the assistant turns it into a safe, read-only query against the live database, applies the same logic our reports use, and replies. It must work by text and voice, in English and Spanish, and a customer must only ever see their own data. What it must do • Take a natural-language question and translate it into a safe, read-only, validated query against the live CattleXpert database (no writes, allow-listed tables, mandatory tenant filter). • Return real-time results, and reproduce existing report logic (e.g. head on feed, ADG, days on feed, dry matter intake, death loss, cost of gain) so answers reconcile with our reports. • Support text and voice — speech-to-text in, text-to-speech out. • Work in English and Spanish for both text and voice, answering in the language asked. • Enforce strict per-customer data isolation, with an audit log of every question and the query it generated. • Be deliverable as a service plus a chat/voice UI that embeds into the CattleXpert 2.5 web app (Node.js / React). What we'll provide • a) A full database with all relevant data (or a representative copy). • b) Access to the codebase to understand how reports and data are produced. • c) Subject-matter knowledge on the CattleXpert application. • d) A set of sample questions with verified answers — also used as the accuracy benchmark. Deliverables • Working CX AI assistant (text + voice, English + Spanish) deployable to our environment and embeddable in CX 2.5. • Source code in our Azure DevOps / Git, plus setup, deployment, and admin documentation. • Test results against the sample Q&A benchmark. Acceptance • ≥ 90% correct on the provided sample Q&A, reconciling to the matching reports. • No query can read another customer's data, write data, or bypass the validation layer. • Equivalent English/Spanish and text/voice questions return equivalent, correct answers in real time. Engagement Fixed bid, milestone-based. Suggested milestones: (1) discovery + query/semantic layer, (2) core English text engine passing the benchmark, (3) report logic + Spanish, (4) voice, (5) security hardening, CX 2.5 integration, UAT and handover. Your proposal should include • A single fixed total price with a per-milestone breakdown, and a timeline. • Your proposed stack and approach to accuracy, query safety, multi-tenancy, and bilingual voice. • Relevant prior “talk-to-your-data” and/or bilingual voice work, plus any questions or assumptions.