Skip to content

solution-architecture.md

Latest commit

 

History

History
80 lines (62 loc) · 2.4 KB

solution-architecture.md

File metadata and controls

80 lines (62 loc) · 2.4 KB

AegisOps Solution Architecture

Goal

AegisOps turns multimodal incident evidence into a reviewable incident report without exposing provider secrets to the browser.

System boundary

  • React/Vite UI
    • collects screenshots, logs, and operator context
    • renders runtime posture, summary pack, and replay proof
  • Express API
    • validates payloads
    • owns model keys
    • normalizes report output
    • exposes dashboards
  • Model runtime
    • Gemini live mode
    • Ollama local mode
    • static/demo fallback mode
  • Export boundary
    • JSON
    • Markdown
    • Slack/Jira
    • optional Google Workspace artifacts

Deployment topology

flowchart LR
  Operator[Operator Browser] --> UI[React UI]
  UI --> API[Express API]
  API --> Provider[Gemini or Ollama]
  API --> Eval[Replay Eval Summary]
  API --> Export[Slack Jira Workspace Exports]
  API --> Meta[Health Meta Summary Pack Schema]
Loading

Reliability posture

  • Browser never owns the Gemini API key.
  • Review surfaces are available even when the live model is unavailable.
  • Replay evals provide a deterministic quality floor before a confidences the live path.
  • Report schema is explicit, so downstream exports do not depend on loose model formatting.

Security and trust boundary

  • provider keys stay on the server
  • multimodal uploads are bounded and validated before model invocation
  • runtime modes are explicit: demo, Gemini live, Ollama local
  • export routes are separated from incident reasoning routes

Operator flow

  1. Open /api/healthz to confirm runtime mode.
  2. Open /api/summary-pack to inspect replay proof and trust boundary.
  3. Open /api/schema/report to verify the contract.
  4. Run one incident path through analyze -> export -> follow-up.

What makes this useful for an AI engineer

  • multimodal input normalization
  • structured output stabilization
  • fallback modes
  • evaluation-backed quality claims
  • explicit contract surface for downstream systems

What makes this useful for a solutions architect

  • clear browser/server/provider boundary
  • review-first API surface
  • export and integration boundary separated from inference boundary
  • production hardening path is visible without reading the whole codebase

Production hardening next steps

  • add request-level audit persistence for exports
  • add SSO/RBAC around incident visibility
  • add deployment-specific env profiles and IaC
  • add latency and cost scorecard per runtime mode