How it works

Qorpera is built on a single architectural bet: memory, not reasoning, is the moat. Every frontier LLM accesses the same models. What no one else has built is the structured, durable, self-updating organisational memory that makes those models operationally useful for a specific business.

Four loosely-coupled layers compose the system.

1. Ingestion

Connectors sync data from your business tools on configurable intervals. Each connector yields a discriminated union of event, content, and activity. Events (HubSpot, Stripe) become entities. Content (communications, documents) is chunked, embedded into pgvector, and indexed for retrieval. Activity signals (email threads, Slack messages, calendar events) feed timelines on person and department wiki pages.

2. The wiki

All operational state lives on wiki pages. A KnowledgePage is a coherent article with structured properties (JSONB), full-text search (Postgres FTS with GIN index), and versioned trust levels. Person profiles, domain hubs, situation instances, project deliverables — every object that matters is a page. The reasoning engine reads pages; the execution engine writes results back to pages.

3. Situation detection

Two detection paths run continuously. Property-based detection scans entity state on a cron (structured rules, natural language, or hybrid). Content-based detection evaluates incoming communications in batches — the system reads emails, Slack messages, and Teams threads, catching commitment language, escalation patterns, and action items no structured rule could catch.

4. Reasoning and action

Every detected situation creates a situation wiki page and dispatches the reasoning engine. Context is assembled from the wiki, activity timelines, communication excerpts (pgvector retrieval), and cross-department signals. Policy rules are evaluated before reasoning begins. The output is a structured action plan proposed for human approval.

Closed-world reasoning is load-bearing. If the AI cannot justify an action with concrete evidence from your systems, it flags missing context — instead of improvising. Citations are mandatory.

Self-improving loops

Every approval, edit, and rejection feeds back. Resolved situations distil raw learnings into heuristics on the situation-type playbook page. Approval rates per situation type surface in governance telemetry, making the delegation decision data-driven. Detection sensitivity adjusts over time; detection drift triggers audit.