Kanvaz AI Labs

Technology

Better forecasts don't fix it. Causal models do.

The bullwhip is structural. We model the structure, find the cause, and let planners act on it every day.

01 — Causal model

Four drivers. One amplification chain.

  1. 01

    Batching

    Orders placed in lumps, not in line with consumption.

  2. 02

    Promotions

    Demand spikes read as trend and pushed upstream.

  3. 03

    Shortage gaming

    Over-ordering when supply looks tight.

  4. 04

    Broken data

    Each tier sees a different version of the truth.

causal_model.scm — tap a nodeIllustrative weights
β 0.34β 0.27β 0.18β 0.21β 0.62β 0.48PromotionsOrder batchingShortage gamingSignal latencyOrder varianceExcess stockStockouts

Node

Order variance

The measured bullwhip: variance amplified per tier.

do(intervention) → simulate effect before changing a parameter.

P(Y | do(X)) ≠ P(Y | X)

Intervention, not correlation

We estimate what happens when you change a policy — not just what moved together in the past.

Vᵢ := fᵢ(PAᵢ, Uᵢ)

Structural causal model

Each tier is modelled as a function of its upstream drivers plus noise, in a directed acyclic graph.

t_valid × t_recorded

Bi-temporal record

GlossDB keeps what was true and when it was known, so every decision can be replayed and audited.

Research brief · v0.1

Causal attribution of order-variance amplification

§1Problem
Order variance amplifies upstream: Var(Oₖ) / Var(D) grows with tier k. Forecast accuracy alone does not reduce it.
§2Model
Each tier is a structural equation Oₖ := fₖ(Fₖ, Bₖ, Pₖ, Lₖ, Uₖ) — forecast, batching, promotion, lead-time and noise — in a DAG.
§3Identification
Policy effects are estimated as P(Oₖ | do(Bₖ = b)) using back-door adjustment on observed confounders, not raw correlation.
§4Validation
Every estimate is replayed on the bi-temporal record: what the model knew at time t, versus what happened after.

02 — Operating loop

A daily loop, not a one-off project.

  1. Step 1

    Sense

    Measure amplification tier by tier.

  2. Step 2

    Explain

    Attribute it to its causes.

  3. Step 3

    Decide

    Approve the right intervention.

  4. Step 4

    Learn

    Measure the effect. Repeat daily.

planner_console · today's exceptionsIllustrative
  • EX-2041Lager 6×33cl · DACHPromo uplift misread as trend → −18% excess stockconf 0.91
  • EX-2042Oat drink 1L · NordicsBatching at DC level → −11% order varianceconf 0.84
  • EX-2043Snack mix 200g · BeneluxStale POS feed (3 days) → Data load quarantinedconf 0.97

03 — Trust

Built for enterprise security review.

Deployment
Runs in your cloud account or data centre.
Data
Read-only access. Your data stays on your infrastructure.
Models
Your data is not used to train shared models.
Audit
Every recommendation and approval is logged and replayable.
Identity
Single sign-on through your identity provider.
Company
Swiss company, engineering in the Zürich region.
Book a technical briefing