Kanvaz AI Labs

Research

Causal science for
supply chain decisions.

We study why order variance amplifies upstream — and how to remove it with interventions planners can trust.

Research areas

P(Y | do(X))

Causal inference

Identifying which drivers create order variance — and what happens when a policy changes.

t_valid × t_recorded

Semantic data

Bi-temporal, governed definitions so every tier reasons over the same version of the truth.

propose → approve → learn

Human–agent decisions

How agents propose, planners approve, and every decision stays auditable.

Working briefs

  • Causal inference

    Structural interventions in multi-tier supply networks

    A structural causal model of order amplification across retailer, distributor, manufacturer and supplier, and how batching, promotions and shortage gaming can be separated from true demand.

    In preparation
  • Semantic data

    Bi-temporal replay for planning decisions

    Reconstructing what a planner knew at decision time versus what was true, so forecast error and decision quality can be evaluated fairly.

    In preparation
  • Causal inference

    Counterfactual evaluation of batch-size policies

    Estimating the variance reduction of smaller, more frequent orders before changing anything in production.

    In preparation
  • Human–agent decisions

    Governance levels for planning agents

    A permission model for when agents may only flag, may propose, or may act within agreed limits.

    In preparation

Patents

Patent pending

Safe causal learning for supply chain decisions

Operational data rarely shows what truly causes what. We are developing a method that learns cause and effect through small, carefully bounded changes in day-to-day operations, always within safe limits, and keeps a clear record of every step.

Causal, not correlational

Identified causal graphs that answer “what if we change X?”

Safe by construction

No experiment runs unless it stays inside the combined safety envelope.

Auditable

Deterministic, replayable records of every decision.