Bilingual policy and citizen-service assistant
One trusted source of answers for front-line staff and citizens, grounded in laws, circulars and service manuals in Arabic and English.
The problem
Staff search across circulars, PDFs and legacy portals to answer the same questions. Answers vary between offices, Arabic and English versions drift apart, and the data cannot leave the country.
How it works
- Ingest circulars, service manuals and FAQs in Arabic and English, with OCR for scanned documents and version tracking for every policy.
- Hybrid keyword and semantic search, filtered by each department's access rights.
- An open-weight, Arabic-capable language model on sovereign infrastructure drafts the answer and cites the exact clause it relies on.
- Sensitive or low-confidence questions are routed to an officer. Every question, answer and source is logged.
Reference architecture
KPIs we track
- Average handling time per enquiry
- First-contact resolution rate
- Share of answers with a valid citation
- Escalation rate to officers
- Citizen satisfaction score
Governance and risk controls
- Data stays in-country, with no calls to external AI APIs
- Human review for any answer that interprets policy
- Audit logs retained to the entity's records policy
- Red-team testing on Arabic and English prompts before launch
How TCO stays controlled
- Fixed-capacity private inference instead of per-token fees at citizen scale
- One shared retrieval core reused across departments and ministries
- A small model routes questions; the larger model only drafts answers
Indicative delivery path
- Validate2 weeks: data audit and a 50-question test set
- Prototype4–6 weeks: one service domain
- Pilot8–12 weeks: one front office with live users
- ScaleMore domains, then the citizen channel
Built on NexusAI (opens in a new tab)