Vision

Capture fine-grained user activity telemetry (apps, windows, timestamps) and convert it into sessions and business intent. Use this to reveal process friction, rework, and automation opportunities—while keeping privacy and governance as first-class constraints.

• Outputs: dashboards, process maps/heatmaps, automation opportunity rankings.
• Success: high intent/session accuracy, measurable reduction in context switching and friction.

Architecture

• Endpoint: Telemetry Agent → Local aggregation → Session pre-processing.
• Ingestion: Secure API + deterministic schema enforcement → raw + session stores.
• Intelligence: Sessionisation rules/DSL → activity classification → intent inference.
• Consumption: KPI engine → dashboards/APIs → agentic applications.

The system-design page shows the same flow as a color-coded SVG.

Semantic Layer

This is the organisation’s “process intelligence core”—the governed layer that turns telemetry into reusable knowledge.

• Reference DB: standards, subprocess maps, known activities.
• Correlation DB: patterns, signatures, cross-user sequences.
• Feedback DB: human validation + agent feedback + rule adjustment logs.
• Enterprise memory: knowledge graph + process intelligence store for reuse at scale.

LLM Philosophy

LLMs can accelerate discovery and explanation, but must not become the source of truth. Anything compliance-grade must be reproducible and auditable.

• Approach 1: LLM outputs are provisional and must be promoted into deterministic rules/ML before they are authoritative.
• Approach 2: deterministic-first; LLMs are optional and never define truth.

Agents

Agents are guard-railed: they observe, recommend, and learn under governance—never self-authorize changes.

• Observation Agent: monitors workflows and KPI changes.
• Recommendation Agent: proposes next-best actions / automation candidates.
• Learning + Audit: improvements are logged, reviewed, and promoted via governance gates.

Developers

• Keep ingestion + event schemas deterministic for replayability and testing.
• Version rules/DSL; add observability (logs, traces, audit).
• Use LLMs only as non-authoritative helpers (summaries/naming) with expiry + promotion workflow.

Delivery Plan page includes sprint-by-sprint ownership and a readable Gantt chart.