Architecture

How Capped AI deploys an isolated inference environment inside your AWS VPC.

Every component in the inference path runs on a physical appliance inside your plant perimeter. No operator data leaves the site. Capped AI connects to your DCS historian in read-only mode via OPC-UA. The appliance cannot write to or actuate the DCS — architectural guarantee, not policy.

System architecture

┌─────────────────────────────────────────────────────────┐
│                   Plant Perimeter                        │
│                                                          │
│  ┌──────────────┐    OPC-UA     ┌─────────────────────┐ │
│  │  DCS / DeltaV │   (read-only) │  Capped AI Appliance    │ │
│  │  Honeywell   │──────────────▶│  1U · 2× H100 80GB  │ │
│  │  PI Historian│               │  Qwen 2.5 72B        │ │
│  └──────────────┘               │  Whisper-large-v3    │ │
│                                 │  pgvector retrieval  │ │
│  ┌──────────────┐               └──────────┬──────────┘ │
│  │  Operator    │ Voice/text               │            │
│  │  Tablet      │──────────────────────────┘            │
│  └──────────────┘                                        │
└──────────────────────────────┬──────────────────────────┘
                               │ Outbound only
                               │ Aggregate metrics, no customer data
                               ▼
                    ┌──────────────────────┐
                    │  Capped AI Ops Cloud     │
                    │  Health metrics only │
                    │  Model updates only  │
                    └──────────────────────┘

Key architectural properties

PropertyHow it works
No cloud egress of customer dataAll AI inference runs on the local H100 appliance. Handover text, alarm logs, and DCS readings never leave the plant perimeter.
OPC-UA read-onlyCapped AI connects to your historian via a read-only OPC-UA subscription. The appliance has no write path to the DCS — verified at the firewall and at the OPC-UA session level.
Open-weights LLM on-siteQwen 2.5 72B runs in vLLM on the H100 appliance. Llama 3.3 70B is available as a fallback. No dependency on external AI APIs.
WORM audit logEvery completed and signed handover is written to a tamper-evident, append-only log on the appliance. Cannot be deleted without physical access and a dual-person integrity process.
Voice to text on-siteWhisper-large-v3 runs locally for Arabic and English speech recognition. No audio data is sent off-site.
pgvector retrievalHistorical handovers are indexed in pgvector for semantic search. The Search module queries this index; the data stays on the appliance.

DCS integrations — reference architectures

Select your DCS vendor below. Each section includes the network topology, data-flow table, and a configuration checklist for your plant IT team.

Capped AI connects to Honeywell Experion PKS and Uniformance PHD via the built-in OPC-UA server available from Experion R500 onward. No third-party middleware is required — the appliance opens a read-only OPC-UA subscription directly to the Experion server on port 4840.

PLANT PERIMETERHoneywell ExperionPKS / Uniformance PHDOPC-UA ServerTCP port 4840OPC-UA TCP 4840read-only · no reverse pathCapped AI Appliance1U · 2× NVIDIA H100 80GBvLLM · Whisper-large-v3 · pgvector✓ No write path to DCS✓ Inference runs on-site only✓ Handover data never leaves siteUbuntu 22.04 · air-gapped by defaultHTTPS outbound · metrics onlyCapped AI Ops Cloudaggregate health metrics · no customer data

Data flow

Data typeHistorian sourceTypical latency
Process values (analog / discrete)Experion PKS real-time DB50–150 ms
Alarm / Event logExperion Alarm Manager50–150 ms
Sequence of Events (SoE)Experion Event Journal50–150 ms
Operator journal entriesExperion Station notesOn-demand
Historical trend dataUniformance PHD (up to 5 yr)100–300 ms

Configuration checklist

  • Experion R500+ (PKS 510 / 520 / 530) with the built-in OPC-UA server enabled
  • OPC-UA endpoint URL: opc.tcp://<dcs-server>:4840 (or custom port configured in Experion)
  • Security policy: Basic256Sha256; message mode: SignAndEncrypt
  • Generate Capped AI read-only OPC-UA client certificate (2048-bit RSA, 2-year expiry) — provided by Capped AI field engineer
  • Add Capped AI client cert to Experion's Trusted Certificates store in the Experion Server Configuration Console
  • Grant NodeId read access to: AlarmConditionType (alarm summary), historical trend nodes, and Operator Journal — no write-capable NodeIds
  • Firewall rule: allow TCP 4840 from Capped AI appliance IP to Experion OPC-UA server IP (one-way, inbound to DCS only)
  • Verify no write-capable NodeIds appear in the Capped AI session's subscription list after connection

Integration FAQ

Does Capped AI modify setpoints?
No. The OPC-UA connection is established as a read-only Subscribe session. Capped AI holds no write credentials and no write path exists at the protocol level. This is verified jointly by the Capped AI field engineer and your DCS engineer at installation.
Can Capped AI trigger alarms?
No. Capped AI reads the alarm and event log from your DCS historian. It cannot write to the DCS, cannot create or acknowledge alarms, and cannot send any control signals. Firewall rules document this at the network layer: no inbound path from the Capped AI appliance to the DCS write interface.
What ports does Capped AI use?
TCP 4840 — OPC-UA, outbound from Capped AI appliance to DCS historian (inbound to DCS only; no reverse write path). TCP 443 — HTTPS, outbound from Capped AI appliance to ops.capped.ai for the encrypted aggregate metrics heartbeat (no customer data; can be disabled for fully air-gapped mode). TCP 443 — HTTPS, inbound from operator tablets to Capped AI appliance. No other ports are opened or accepted.
Can Capped AI operate fully offline with no internet connection?
Yes. The outbound heartbeat to Capped AI Ops Cloud can be disabled on request. All AI inference (Qwen 2.5 72B in vLLM), speech recognition (Whisper-large-v3), and handover data storage (pgvector on PostgreSQL) run on the local appliance. The appliance is fully air-gapped by default during production — the heartbeat only carries aggregate health metrics, never customer data.

Appliance hardware specification

ComponentSpecification
Form factor1U rack-mount server
CPUIntel Xeon Gold 6438N (32-core, 2.0 GHz)
GPU2× NVIDIA H100 80GB NVLink (SXM5)
RAM256 GB DDR5 ECC
Storage8 TB NVMe RAID-1 (handover data) + 2 TB OS SSD
Network2× 10GbE (OPC-UA + management), 1× IPMI
Power220V / 15A, redundant PSU
Operating systemUbuntu 22.04 LTS (hardened, no public internet access)
InferencevLLM serving Qwen 2.5 72B + Whisper-large-v3
DatabasePostgreSQL 16 + pgvector extension