The platform is built the way a national service has to be built: no single machine that can fail, capacity that grows automatically when demand grows, an edge that absorbs attacks before they reach the application, and every server, disk and AI model inside Qatar's borders. The same design runs in a sovereign cloud region, on NCSA's own hardware, or fully disconnected.
Labels are deliberately plain. The technical specification (instance classes, GPU counts, network ranges, HSM model, exact failover timings) is agreed with NCSA's architecture team during the proof of value and recorded in the signed design document.
Pages and images are served from points of presence in Doha so the platform feels instant anywhere in Qatar. The same edge runs Spark SHIELD: it blocks attacks, bots and floods before they reach a server.
Every request is handed to a healthy application server. If a server fails, the balancer stops using it within seconds and nobody notices. Runs across two availability zones itself.
Application servers are identical and stateless. When CPU or request queues rise past an agreed threshold, more servers are started automatically; when load falls, they are removed. Capacity follows demand, not a guess.
The database writes to a primary and a synchronous replica in a second zone. Evidence files live in object storage replicated the same way. Keys are held by NCSA in an in-country HSM.
Open-weight models run on GPU nodes in the same boundary. No prompt, record or finding is sent to any outside AI service. More GPU nodes are added behind the balancer as the number of analysts grows.
Encrypted backups go to a second in-country site hourly. A full restore is rehearsed every quarter with NCSA present, and the measured recovery times are recorded in the audit log.
| Tier | Scales on | How | Agreed limits |
|---|---|---|---|
| Edge | Request volume, attack traffic | Provider capacity in Doha points of presence; no action needed | Unlimited within the in-country edge |
| Application | CPU above 60 % for 3 minutes, or request queue depth | New identical nodes join the balancer automatically; removed after 15 quiet minutes | Minimum 2 nodes (one per zone), maximum set per contract |
| Scheduled scanning | Number of authorised entities and assets | Worker nodes scale separately from the web tier so a large scan never slows the console | Sized from the connected estate |
| Sovereign AI (SRE-1) | Concurrent analysts, answer latency | Additional GPU nodes added behind the balancer; model weights pre-staged | Model tier and node count agreed with NCSA |
| Data | Storage used, query latency | Storage grows online; read replicas added for reporting | Retention 7 years by default |
