Architecture overview.
Vavoris is built on ten platform layers that together form a closed loop: signals in, decisions out, outcomes measured, system improved. This page describes each layer and how they interconnect — for technical evaluators, architects, and integration teams.
The six products.
Each product maps to a distinct phase of the decision-to-outcome loop. Together they form a complete closed system.
Connect™
15 native source connectors cover every common enterprise transport — streaming platforms, cloud storage, relational and columnar databases, and API endpoints. Field mappings are stored centrally and applied automatically at ingest time, with zero custom code required per source. New sources go live without system restart.
Observe™
Signal history accumulates continuously in isolated, per-tenant state. Every historical score is weighted by recency — recent signals carry more weight than older ones. No two tenants share state at any layer. The accumulation layer is what makes every subsequent decision context-aware rather than reactive.
Decide™
A composable rule graph engine evaluates each signal against a purpose-built decision graph for the use case. Each graph is configured independently — combining unrelated scoring logic dilutes both. Rule graphs reload instantly without downtime. An Outcome-Aware Guardrail blocks recommendations when the graph's measured precision falls below the configured threshold, and passes through gracefully when no precision history exists yet.
Govern™
Four governance modes: Recommend (suggestion surfaces, human decides), Human Approval (every decision requires a designated approver before action), Hybrid Review (human reviews and edits the recommendation before approving — every edit captured as calibration data), and Guarded Automation (automated within pre-approved policy scope with configurable spending limits; any action exceeding governance thresholds is escalated, not silently executed). Every decision and governance event is recorded in a tamper-detectable, append-only audit chain.
Act™
Delivers recommendations and governed automated actions through webhook, email digest, and push notification channels. A quality feedback loop monitors advisory effectiveness continuously — sessions that fall below quality thresholds are flagged and quarantined automatically.
Learn™
Nightly calibration adjusts recommendation thresholds based on measured precision against actual outcomes. An automated shadow evaluation system tests candidate decision graphs against live traffic and promotes better graphs when they demonstrably outperform the current one. Model retraining uses outcome labels as ground truth — accuracy shifts from heuristic scoring to supervised learning as outcome coverage grows. Every calibration event is auditable: the platform explains what changed and why.
Key architectural properties.
These properties hold across all platform layers and all tenants. They are enforced by design, not convention.
- Scoring state is fully isolated per tenant — no cross-tenant data sharing at any stage of processing.
- Decisions are immutable once produced. The audit chain is append-only and tamper-detectable.
- Rule graph configuration changes propagate to all active pipelines instantly, with zero downtime. Running decisions complete under the prior configuration; new decisions use the updated one.
- The Outcome-Aware Guardrail blocks recommendations when a graph's measured precision falls below the configured minimum. It passes through gracefully for new graphs that have not yet accumulated outcome history.
- Every decision includes a full explanation — which signals contributed, how each scored, and what the recommendation would have been under the prior graph configuration (counterfactual score). When recommendation behavior changes over time, the platform generates a plain-language why-changed narrative explaining what changed, when, and by how much.
- Guarded Automation enforces configurable spending limits before any autonomous action. No limit applies when no cap is configured.
- Tenant isolation is enforced independently at every platform layer — processing state, operational storage, and outcome records are all scoped to the individual tenant.
The outcome loop.
Every decision produces an outcome. Every outcome improves the next decision. The loop is automatic — no manual intervention required for the core calibration path.
When a decision resolves — through human approval, automated execution, or timeout — the platform joins the decision record with the outcome and stores the pair in a per-tenant outcome index. Three calibration processes read from this index nightly.
Threshold calibration computes the precision of each decision graph against actual outcomes and adjusts recommendation thresholds proportionally when precision falls below the configured target. Adjustments propagate to all active pipelines before the next day's traffic — no deployment required.
Shadow graph promotion runs candidate decision graphs in parallel against live traffic without affecting production recommendations. When a shadow graph demonstrably outperforms the current live graph — by a meaningful margin with sufficient validated samples — it is promoted automatically. No human intervention required.
Model retraining uses outcome labels as ground truth — the highest-quality training signal available. When outcomes are not yet available for a decision, the platform fills the gap with other signals and shifts toward supervised learning as outcome coverage grows. The model learns from what actually worked, not from what the scoring function predicted.
The result is compounding accuracy: each nightly cycle produces a decision graph and model calibrated to one more period of actual outcomes in the tenant's specific operational context. The longer the platform runs, the more its recommendations reflect what works — in that domain, with that data, under those conditions.