Verified Bayesian Decision Science

DATA INTELLIGENCE VERIFIED

Bayesian models for marketing, customer and forecasting decisions. Every result is checked by five verification layers and approved by a person before anyone acts on it.

5
Verification Layers
12
Decision Packs
LAUNCH APP →

Decision Engines

Four engine classes — forecasting, causal experimentation, marketing measurement, and customer economics — across all 12 decision packs today, run on demand or on a schedule you set.

See all 12 packs
Forecast

Forecasting & Monitoring

Packs Demand · Time series · Anomaly
Causal

Causal & Experimentation

Packs Causal impact · A/B testing
Measurement

Marketing Measurement

Packs MMM · MTA · Synthetic control · Geo
Retention

Customer Economics

Packs LTV · Churn · Elasticity

New packs slot into these engine classes and run through the same verification stack. How verification works →

How TriaCaf works

Watch. Re-check. Recommend.

This is the loop TriaCaf is built around: watch the numbers you care about, re-check the models when something moves, and bring a verified recommendation to a person. The aim is for all three steps to run on their own; that is what we mean by always-on. It is rolling out in stages, so each step below says what it will do and what runs today.

Live today

Run any of 12 Bayesian packs on demand or on a schedule; every result verified and human-approved.

  1. 01 · Watch Rolling out

    Keep an eye on the metrics you choose as new data arrives, and notice when one moves outside the range its model expects.

    Today: anomaly checks on your latest upload, run on a schedule you set.

  2. 02 · Re-check Rolling out

    When something moves, re-run the packs that explain it on the new data, through the same five verification layers as any other run.

    Today: you start a re-run on demand, or put it on a schedule.

  3. 03 · Recommend Rolling out

    Turn a verified result into a recommendation with its confidence attached, for a person to approve or send back.

    Today: every run pauses for a person to approve it, and approved runs produce a signed Decision Brief.

What it's built with

PostgreSQL
PyMC
ArviZ
Modal
Parquet
pandas
NumPy
PostgreSQL
PyMC
ArviZ
Modal
Parquet
pandas
NumPy

Our
Expertise

Bayesian models across 12 decision packs, with the checking built in: every run is fit more than one way, verified layer by layer, and approved by a person before anyone acts on it.

Also: customer economics, forecasting, causal inference and A/B testing. All 12 packs →

  1. 01Data

    You upload a CSV export.

  2. 02Profile

    Columns typed; gaps and outliers flagged before any fit.

  3. 03Model specs

    Two or three model specifications, each fit on its own.

  4. 04Verification

    Execution, diagnostics, predictive check, sensitivity.

  5. 05Human approval

    The fifth layer: a person approves or rejects the result.

  6. 06Decision Brief

    PDF or slide, carrying the run's Ed25519-signed model card.

The same path for all 12 packs

How a run works

Upload a CSV and TriaCaf profiles it, fits two or three model specifications, checks them layer by layer, and waits for a person to approve before producing a signed Decision Brief.

Business Solutions

Applied Intelligence.

Four questions teams bring to TriaCaf most often, and the decision packs that answer them. Each links to the live use case on the Solutions page.

01 //

Customer Lifetime Value (CLV) Prediction

Estimate the long-term value of each customer cohort, with credible intervals, so acquisition spend can be judged on long-run value rather than first-order conversions.

See the use case
02 //

Demand Forecasting & Monitoring

Hierarchical forecasting paired with anomaly checks that run on a schedule you set, so a real shift in demand, spend or usage is flagged without waiting for someone to open a dashboard.

See the use case
03 //

Causal Inference & Experimentation

Quasi-experimental lift estimation and Bayesian A/B testing for the decisions that can't run a clean randomized trial — with a real counterfactual, not a correlation.

See the use case
04 //

Marketing Mix & Budget Allocation

Bayesian MMM quantifies the incremental revenue driven by every channel, giving finance and marketing a shared, auditable basis for reallocating budget.

See the use case
The Paradigm Shift

Why the Old Model is Broken.

Most analytics reports what happened. TriaCaf estimates what drove it, says how sure it is, and has a person sign off before anyone acts.

Traditional Analytics
The TriaCaf System

Manual Spreadsheets

Data is downloaded weekly into disconnected Excel files, causing reporting delays and human error.

One Validated Dataset

Upload your exports once; every dataset is profiled and validated before a model ever sees it.

Weekly BI Dashboards

Someone has to remember to open the dashboard. Nothing is actually watching for a real shift as it happens.

Scheduled Forecasting & Anomaly Checks

Hierarchical forecasts and residual-based anomaly checks run on a schedule you set, so a real deviation is flagged without anyone having to remember to look.

Retrospective Analysis

Looking backwards at what happened last month to guess at what should happen next month.

Forward-Looking Bayesian Models

Bayesian models estimate lifetime value, churn risk and demand with credible intervals, re-run on demand or on a schedule you set as new data comes in.

Correlation Presented as Proof

A metric moved after the change shipped, so the change gets the credit — no counterfactual, no rigor.

Verified Causal Inference, Human-Approved

Quasi-experimental and Bayesian methods estimate what would have happened anyway, so credit only goes where it's actually earned — reviewed by your team before anything moves.

Onboarding & verification

From first upload to a signed decision.

How a new team gets started, and the checks every result passes before it reaches the people who act on it.

01
Phase 01

Scope the decision

We start from the decision you need to make, pick the decision packs that answer it, and agree what data each one needs.

02
Phase 02

Data Onboarding

Mapping your exports into clean, profiled datasets — the dates, channels, outcomes and covariates each decision pack needs — validated before anything is fit.

03
Phase 03

Modeling & verification

Running the decision packs that fit the question — Bayesian models fit across multiple specifications, each result checked for convergence, fit, and cross-spec agreement.

04
Phase 04

Human approval & Decision Brief

A person on your team approves or rejects every result. Approved runs produce a Decision Brief with a signed model card, and you can schedule re-runs to refresh the answer as you upload new data.

Latest Insights

How TriaCaf checks a result before anyone acts on it, and how to read what it hands you.

Verification
6 min read

Five checks before you see a number

What each of TriaCaf's five verification layers checks on every run, and why the last one is a person.

Read article
Page one of the sample Decision Brief, built from demo data
Sample brief · demo data
5 min read

How to read a Decision Brief

The verdict banner, metric tiles, verification strip, caveat and signed model card, section by section.

Read article
Methodology
6 min read

Bayesian MMM without a black box

Adstock, saturation, three specifications and ROAS ranges: how the marketing mix modeling pack works.

Read article

The Intelligence Stack

Open-source Bayesian tooling you can audit, running on managed cloud compute. Data comes in as CSV today; Google Ads and Meta connectors are in development.

PyMC
PyMC-Marketing
SciPy
Modal
PostgreSQL
Railway
Python
PyMC
PyMC-Marketing
SciPy
Modal
PostgreSQL
Railway
Python
Free Diagnostic

What is Your Data Maturity Score?

Are your decisions based on lagging indicators or on models you can check? Six questions place you on the same four-rung ladder we use to describe the platform: reporting, monitoring, modeling, and verified decisions.

  • Identify the gaps between what you're measuring and what's actually driving the outcome.
  • See which rung your team works from today, and what the next one up would take.
  • If it's useful, book a call afterwards to talk through your result.

Your answers stay in your browser. Nothing is sent.