A decision engine
built to grow revenue and profit.

Kapnova combines your data, outside signals, business rules, economics and operating constraints to find the decisions that matter, then uses purpose-built quantitative models to estimate what each one is worth.

One engine. Every decision makes the next one better.

The surfaces you touch.

Decision simulator

Ask in plain language. Get a quant-grade answer.

A pricing, promotion, loyalty, or media question, answered with the drivers, the confidence, and every number traceable to a source.

DECISION SIMULATORplain-language
Raise Hydra-Glow price 8%?
−$0.4M80% CI [−$0.9M, +$0.1M]Hold
Price elasticity
+4.2%
Sentiment backlash
−6.3%
Net causal effect
−2.1%
4 sourcesidentification: quasi-experimentalevery number traceable
DECISION BRIEF / hydra-glow / pricingone page
Causal effect
−2.1%
net revenue
Correlational
+6.0%
elasticity only
80% interval
−3.4% / −0.6%
10k paths
Recommendation
Hold
test 4% on a smaller SKU
the one thing that would change it: sentiment recovery
Decision brief

One page you defend to finance.

The action, the dollar impact against the correlational read, the confidence, the one thing that would change the answer, and how much each driver contributed.

The assets that accumulate.

Decision ledger

A scoreboard for the budget room.

Every call logged with its prediction and, when the outcome lands, captured. A running total of dollars gained or protected against the correlational read, your number to carry into the budget conversation.

DECISION LEDGER+$2.4M vs correlation
DecisionCallOutcomevs correlation
Hydra-Glow +8%HoldHeld+$1.6M
Q3 creator pushScaleScaled+$0.9M
Loyalty tier expansionNarrowNarrowed+$0.2M
Spring promo cadencePull backPendingtracking
every call scored against the correlational read
DECISION MONITORbetween decisions
Sentiment / 30d
Drivers
Steady
Sentiment
Souring ↓
Transactions
Holding
Flag. Sentiment is souring ahead of transactions. Revisit Q3 pricing before it reaches the P&L.
Decision monitor

It watches between decisions.

It tracks the drivers and the sentiment signal and flags when the picture shifts, sentiment souring ahead of transactions, so you can revisit a call before it reaches the P&L.

Walk into the room with the evidence behind you.

Every driver quantified. Every answer traceable to the data, method and assumptions behind it. And when the evidence isn’t strong enough, Kapnova tells you what is missing instead of pretending to know.

We recommend. You decide.
Audit trail / one number
Raise price 8%? −2.1% traceable
Source3-year transaction panel and review corpus
MethodQuasi-experimental, geo holdout
AssumptionSentiment as a time-varying mediator
Confidence80% CI [−3.4%, −0.6%]

It gets sharper with every decision, and your data stays yours.

Every engagement starts from calibrated category priors, directionally right on day one, and sharpens into your own numbers as your outcomes return.

  • Your raw data never leaves your tenant and is never shared.
  • Only aggregate, de-identified inferences, with no single account reconstructable, inform the priors that warm-start every engagement.
  • Platform-sourced advertising data is excluded entirely; it serves your decisions and never travels.
Day one / calibrated priorsOver time / your numbers

One causal model, not a set of reports.

YOUR DATA THE CAUSAL MODEL WHAT IT PRODUCES SalesSpendCreatorReviewsCausesSimulationsOutcomes every outcome loops back and sharpens the model
The causal data layer

Your data tells us what happened. We structure it to understand why.

Kapnova transforms your existing data so it can separate what actually drove an outcome from what simply moved with it. That foundation makes every decision, scenario and recommendation more defensible.

The external signal layer

Your business doesn’t move on internal data alone.

Kapnova continuously brings in sentiment, reviews, search, competitors and market signals, then models how they influence your outcomes and when.

You don’t have to provide it. We bring the outside world into the model.

The hardest question is which way the arrow runs.

Sentiment is not a static input. Your decisions move it and it moves revenue back, so which way the arrow runs is itself a causal question. The same variable is a mediator in one decision, a confounder in another, a leading indicator in a third. Collapsing that into one score is the mistake we do not make.

If sentiment is the cause
souring → decline

Souring sentiment is driving a real decline. This is a signal to act on.

Move: intervene now.
If sentiment is a symptom
a price change → souring & decline

A price or product change caused both the decline and the souring. The sentiment is an echo.

Move: ignore the sentiment, fix the cause.

Raw observational data in, a decision you can defend out.

Seven steps run on every question, and they run without anyone starting them. The second one is the one that matters, and it is the step a dashboard, an attribution tool and a language model all skip.

AI builds and runs it. The causal engine makes the call.

Working out what caused what used to take a scarce senior specialist weeks per question, built from scratch every time. Agents pre-trained by our MIT team now do that work in hours, at a scale that used to need a room of PhDs.

Language model
The interface

It reads the signal and takes your question. It never invents the answer.

Causal engine
The number

It runs the identified methods and produces the effect, every step auditable.

Human judgment
What does not commoditize

Which assumption to lean on, and where it breaks. This stays with you.

Reinforcement learning

It learns your decision policy, and sharpens with every outcome.

The optimization engine searches your real decision space, budget, margin and inventory included, for the move that maximizes the goal you set, not the metric that looks good this week. As outcomes come back it learns which actions actually caused them, and updates the policy across the whole organization. It starts from calibrated category priors and sharpens into your own numbers with every decision logged.

We are future proofing your enterprise for agent-made decisions.

When a machine makes the call, what sits beneath it decides whether the call can be trusted. A probabilistic language model is the wrong thing to put there: it reasons from correlation and repeats it confidently, without sources, on exactly the highest-stakes calls agents get handed. Those decisions need a causal layer underneath.

YESTERDAYHuman-madeTODAYAgent-assistedSOONAgent-made
The default substrate
An agent on a probabilistic LLM
agent → LLM → correlation
  • Reasons from correlation, and propagates it at machine speed.
  • Confidently and sourcelessly wrong on the calls that matter most.
  • Burns scarce GPU to get there.
The Kapnova substrate
An agent on the causal engine
agent → causal engine → cause
  • Reasons from cause, identified and defensible.
  • Every answer traces to a source, a method, and an assumption.
  • Runs light, on standard cloud.
The compute argument
Same decision. A fraction of the compute.
Probabilistic LLM
scarce GPU
Kapnova causal engine
standard cloud
Causal and simulation models rather than large generative inference. Less to run, and the answer an agent can actually act on. Efficiency here is not a discount; it is evidence the architecture is right.

Start with your URL. Go deeper with your data.

Start with your URL and Kapnova will surface what it can find from public data alone. Then connect your own data to measure what past decisions actually caused and where the next revenue or profit opportunity may be.