Build University AI Assistant
Budget / Salary$5,000–10,000
TypeFreelance project
LocationRemote
Posted1 hour ago
We need an experienced AI engineer or small team to build a production conversational assistant for a university website in Saudi Arabia. The assistant answers questions from students, applicants, parents, staff and visitors in Arabic and English, by text and voice, 24/7.
Scale: about 200,000 queries/month, 20,000 users, and 2,000+ concurrent sessions at peak.
What you will build
RAG question answering: hybrid retrieval, re-ranking, multilingual embeddings, and metadata filters. Every answer cites an official source page.
Zero fabrication: claim-level grounding, a faithfulness check, abstention when unsure, and deterministic answers for fees, deadlines and admission criteria from structured data.
Automated content ingestion: incremental crawler (JS rendering, change detection), layout-aware PDF and table parsing with OCR, a versioned index with rollback, and updates live within 24 hours.
Agentic layer: multi-agent orchestration on an open-source, self-hosted framework (e.g. LangGraph). MCP-style typed tools for ticketing, contact lookup and visit scheduling, with guardrails and fallback to plain retrieval.
Voice: streaming speech-to-text and text-to-speech with barge-in, under 2 seconds end to end.
Admin portal: dashboards, conversation review and correction workflow, role-based access, audit log, and content approval.
Safety: protection against prompt injection and jailbreaks (OWASP LLM Top 10), PII redaction, and scope control.
Evaluation: an automated test harness (faithfulness, precision, recall) and red-team results.
Delivery: a responsive, RTL-correct widget (WCAG 2.1 AA), Kubernetes deployment, CI/CD, documentation, and full source code handover.
Hard constraints
All data and models must stay inside Saudi Arabia. There are no external API calls at runtime, so no OpenAI or similar. Open-weight Arabic-capable models are self-hosted.
Alignment with Saudi PDPL and NCA controls.
Required experience: production RAG with Arabic, LLM serving, Kubernetes, voice pipelines, and agent frameworks. Please send examples of Arabic NLP or RAG work you have shipped.
To apply, please include: a proposed architecture in a few lines, your timeline, your rate by milestone, and two relevant references. An NDA is required before detailed documents are shared.
Scale: about 200,000 queries/month, 20,000 users, and 2,000+ concurrent sessions at peak.
What you will build
RAG question answering: hybrid retrieval, re-ranking, multilingual embeddings, and metadata filters. Every answer cites an official source page.
Zero fabrication: claim-level grounding, a faithfulness check, abstention when unsure, and deterministic answers for fees, deadlines and admission criteria from structured data.
Automated content ingestion: incremental crawler (JS rendering, change detection), layout-aware PDF and table parsing with OCR, a versioned index with rollback, and updates live within 24 hours.
Agentic layer: multi-agent orchestration on an open-source, self-hosted framework (e.g. LangGraph). MCP-style typed tools for ticketing, contact lookup and visit scheduling, with guardrails and fallback to plain retrieval.
Voice: streaming speech-to-text and text-to-speech with barge-in, under 2 seconds end to end.
Admin portal: dashboards, conversation review and correction workflow, role-based access, audit log, and content approval.
Safety: protection against prompt injection and jailbreaks (OWASP LLM Top 10), PII redaction, and scope control.
Evaluation: an automated test harness (faithfulness, precision, recall) and red-team results.
Delivery: a responsive, RTL-correct widget (WCAG 2.1 AA), Kubernetes deployment, CI/CD, documentation, and full source code handover.
Hard constraints
All data and models must stay inside Saudi Arabia. There are no external API calls at runtime, so no OpenAI or similar. Open-weight Arabic-capable models are self-hosted.
Alignment with Saudi PDPL and NCA controls.
Required experience: production RAG with Arabic, LLM serving, Kubernetes, voice pipelines, and agent frameworks. Please send examples of Arabic NLP or RAG work you have shipped.
To apply, please include: a proposed architecture in a few lines, your timeline, your rate by milestone, and two relevant references. An NDA is required before detailed documents are shared.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.