28 lines
1.5 KiB
Markdown
28 lines
1.5 KiB
Markdown
# ADR 0002: Hybrid AI Architecture for SCADA Platform
|
||
|
||
**Status:** Accepted
|
||
**Date:** 2026-06-19
|
||
|
||
## Context
|
||
|
||
The SCADA platform needs to support both real-time local decision making on edge devices (Raspberry Pi) and complex reasoning / long-term analysis in the cloud. A single model size cannot efficiently satisfy both requirements while keeping egress costs low.
|
||
|
||
## Decision
|
||
|
||
We will use a **hybrid AI architecture**:
|
||
|
||
- **Edge (small models, 1B–9B parameters)**: Run on Raspberry Pi for local signal classification, simple alert triage, OCR/needle gauge interpretation, and immediate control decisions.
|
||
- **Cloud (larger models, 30B+ or frontier models)**: Handle complex reasoning, anomaly detection across time, user intent understanding, report generation, and high-level planning.
|
||
- A small "triage" model on the edge decides whether to escalate a situation to the cloud model or handle it locally.
|
||
|
||
## Consequences
|
||
|
||
- Lower ongoing data egress costs (only rich context is sent during alerts or when requested).
|
||
- Better responsiveness for time-critical local events.
|
||
- Increased system complexity (model routing, context handoff, versioning).
|
||
- Requires careful design of the handoff protocol between edge and cloud models.
|
||
|
||
## Alternatives Considered
|
||
|
||
- Pure cloud architecture: Rejected due to high egress costs and latency on poor cellular connections.
|
||
- Pure on-device large models: Rejected due to hardware constraints on Raspberry Pi and inability to handle very complex reasoning. |