How ARIA autonomously monitors the world around you, classifies what matters, and delivers insights before you need to ask — in under 60 seconds for time-critical events.
The Proactive Intelligence Engine (PIE) operates as a three-gate pipeline that auto-chains through increasingly expensive analysis stages. Each gate acts as a filter — only genuinely significant changes reach the final, costliest stage.
Polls 16 data sources against watermarks. Detects changes.
Gemini Flash-Lite classifies changes as noise or signal.
Claude Sonnet performs deep analysis on significant changes.
Routes actionable insights via push, email, or in-app inbox.
The Context Accumulator polls each source using a watermark system — it remembers the last item seen and only processes changes since then. This ensures zero duplicate processing and O(1) per-source overhead.
Gate 2 uses Gemini Flash-Lite (free tier) to classify each context snapshot. Only significant and urgent changes pass through to Gate 3 for deep analysis.
Routine noise. No action, no logging. Quietly discarded.
Worth logging for pattern detection. Not worth alerting on.
Passes to Gate 3 for deep analysis. May generate an insight.
High priority. Passes to Gate 3 immediately. Time-sensitive delivery.
CAFT is a regex-pattern-based acceleration layer that bypasses the 5-minute batch window for time-critical events. It reduces end-to-end latency from ~5–10 minutes to under 60 seconds — without compromising the safety of Gates 2 and 3.
Always-on safety patterns that never change. Hardcoded triggers for emergency scenarios.
Context-aware patterns regenerated every 10 minutes by Gemini Flash-Lite based on current situation.
| Aspect | Normal Pipeline | Fast-Tracked |
|---|---|---|
| Trigger | 5-minute polling cycle | Immediate on pattern match |
| Source | Context Accumulator (16 sources) | Looki Realtime Poll (~60s) |
| Snapshot | Batched with other changes | Isolated, single-event snapshot |
| Entry Point | Gate 1 → Gate 2 → Gate 3 | Direct to Gate 2 (skip batching) |
| DB Marker | fast_tracked = false |
fast_tracked = true |
| Job Priority | 1–2 (standard) | 3 (high) |
| End-to-End Latency | 5–10 minutes | < 60 seconds |
| Significance Gate | Gemini classifies normally | Same — no bypass |
| Risk Enforcement | Applied in Gate 3 | Same — no bypass |
Every 10 minutes, CAFT rebuilds its pattern library using Gemini Flash-Lite and current context. The patterns adapt to where you are, what you're doing, and what time it is.
Even when an insight is generated, ARIA's actions are bounded by the user's configured risk level. Every insight declares the minimum risk level needed to act on it.
ARIA can observe, analyze, and notify. Cannot take action.
ARIA can draft content and prepare actions for your review.
ARIA can act independently — send messages, create events, post.
Gate 3 detects behavioral patterns over time. These patterns have a lifecycle — they grow with evidence, decay without reinforcement, and can be promoted to permanent core memory.
The system automatically adjusts how aggressive it is based on your feedback over the last 7 days.
Six behavioral pattern categories detected by Gate 3.
