ANALYTICS THAT WORKTruth, verified.

AI-based decision making, built by a forecaster

AI has changed how decisions get made. Has it changed how yours get made?

On the drive in, ask your phone what moved yesterday, what’s at risk this week and what to do first. The answers come from your own data, with their accuracy attached. Your plan for the day is ready before you park.

See what we offer Book a conversation
7:42 a.m. · in the car

Illustration: the numbers are made up, the workflow is real.

The shift

Is your team 5–10× more productive than two years ago?

If not, it isn’t for lack of AI tools. Adding a chatbot to the old process buys a modest gain: controlled studies measure people finishing individual tasks about 14% to 40% faster.1,2 The step change comes from changing the process itself, so leaders stop waiting on reports and ask the data directly.

?

Two years ago

  • A question goes into the data team’s queue
  • The answer comes back days later, as a spreadsheet
  • The forecast is refreshed monthly and trusted by no one
  • The meeting is spent arguing about whose number is right

With analytics that work

  • Ask in plain English, at your desk or from your phone
  • Get the answer in seconds, computed from your data, with its definition
  • Every forecast shows its own track record
  • The meeting starts from the same numbers and gets to the decision

What does waiting cost you?

Move the sliders to your team’s numbers.

1,080

decision-days a year spent waiting on answers. With a connector, each of those questions takes seconds.

Requests × days waiting × 12. Illustrative, not a benchmark.

Why data science projects fail

Most analytics projects never change a decision

The research agrees on why, and it is rarely the model. Projects fail because they solve the wrong problem, the data isn’t ready, nothing gets them into day-to-day work, and nobody trusts the output.

95%

of enterprise generative AI pilots showed no measurable profit impact in MIT’s 2025 study. The authors blame brittle workflows and poor fit with daily work, not the models.3

30%

of generative AI projects, at least, were expected by Gartner to be abandoned after proof of concept by the end of 2025.4

5

root causes RAND found in interviews with 65 data scientists and engineers: the wrong problem, missing data, technology for its own sake, no infrastructure, and problems AI can’t yet solve.5

How we work differently

Start from the decision

We begin with the decision you make every week, not with the dataset or the model, and build only what changes it.

Put the data where the questions are

A custom MCP connector makes your data answerable from Claude on any device, with every metric’s definition and source attached.

Prove it before you rely on it

Every forecast ships with its backtest: how it would have done on your own history, and how far off it was at each horizon.

Keep the builder on the line

You talk to the person who built it, who can explain every number in plain English and fix what breaks.

The proof

This is what every model shows you before you rely on it

The model is rebuilt on a growing window of history and asked to forecast the months it hadn’t seen, using only what was known at the time. Then we show you how far off it was.

0.0%
average error, one month ahead
0 of 24
months within 5%
1 mo3.2%
3 mo4.7%
6 mo6.1%
12 mo8.4%

Synthetic series for illustration. Your model shows your numbers.

ActualForecast, made before the month

What we do

License the forecasting engine on your own dedicated server, or have us build the connectors and dashboards that let your whole team answer their own questions.

Licensing

Model Score engine, with its MCP connector

A forecasting engine that finds, tests and explains models for a business series, such as monthly unit sales, from the indicators that might drive it: economic series, prices, weather, calendar effects.

  • Searches thousands of indicator combinations and keeps the models that hold up.
  • Ranks models on how they would actually have forecast, and on whether their drivers make sense and stay stable, not just on fit.
  • Shows its work for every model: backtest forecasts, error by horizon, drivers and stability.
  • Works inside Claude through the included MCP connector, so analysts can run and question models in plain language.
  • Your own dedicated server. Each client runs on a separate server we host, so your data never sits alongside anyone else’s.

TODO: licensing terms and pricing

Ask about a license
Services

Custom MCP connectors and AI dashboards

We connect your data to Claude and other AI assistants through a custom MCP connector, and pair it with dashboards built around the decisions your team makes. People stop waiting on the data team because they can ask the data directly.

  • Built for your systems: your warehouse, internal APIs, forecasts and KPIs, exposed as tools an AI can use.
  • Every number defined: each metric carries its definition, calculation and source, so answers match what finance reports.
  • Dashboards with an AI assistant built in, plus automated briefs that explain what changed and why.
  • Ask it anything: answers are computed from your data, not guessed.

Example: for a 250-location franchise network, requests to the data team fell from about 30 a month to zero once people could answer their own questions. UNCONFIRMED: confirm the figure and that it can be published

TODO: engagement model and starting price

Scope a project

Experience

Over 20 years building pricing, promotion and forecasting models, hands on

Not a sales team in front of an offshore modeling shop. Analytics That Work is Loren Marti, who has written the code and built the models at some of the largest consumer goods, retail and analytics companies in the country since 2002. UNCONFIRMED: results and roles come from your master resume; confirm the figures before launch

$14M

profit increase at a Fortune 500 food manufacturer from an optimal promoted-pricing algorithm, while cutting trade spend by $20M.

$800K

a year saved at a national sales and marketing agency by building an in-house system that measured every event across 4,000+ stores.

10,000+

regression models in production at that agency, producing automated client insights every week.

30% → 13%

forecast error on a hard segment, in a daily revenue forecast for a 250-location franchise network.

  1. 2025 – now250-location home services franchiseData Science Lead: revenue forecasting, pricing, AI dashboards and an MCP server for the executive team
  2. 2019 – 2024Predictive analytics software companyChief Data Scientist: model building methods, R&D, and training a team of 12 economists and model builders
  3. 2016 – 2019National sales and marketing agencySr. Director of Advanced & Predictive Analytics: built the analytics team and its event measurement system
  4. 2012 – 2016Fortune 500 packaged foods companyDirector of Statistical Forecasting & Customer Analytics for its largest retail customer
  5. 2010 – 2012Retail data and analytics providerDirector of Modeling: forecasting and marketing mix models
  6. 2006 – 2010Strategy consultancy, now part of a global consulting firmSenior Consultant: go-to-market strategy for Fortune 500 companies
  7. 2002 – 2006Fortune 500 food manufacturerManager of Market Research and CRM Modeling for its largest retail customer: pricing and marketing mix models

U.S. Marine Corps veteran. MBA. Built three analytics teams from scratch.

Why our mark is a Hebrew word

In the legend of the Golem of Prague, a clay figure was brought to life by writing אמת, emet, “truth”, on its forehead.

To stop it, its maker erased the first letter, the aleph. What remained was מת, met: “dead.” A model that can’t show its validation is a golem with the aleph rubbed off. It still moves, but nothing real is animating it.

Our job is to keep the proof visible.

We keep the aleph.

Let’s see if your forecasts hold up.

Tell us what you’re trying to predict, or what questions your team keeps sending to the data team. You’ll hear back from Loren directly.

Email [email protected]

TODO: decide on a booking link

Sources

  1. Brynjolfsson, Li and Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025: 14–15% more issues resolved per hour.
  2. Noy and Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence,” Science, 2023: 40% less time on writing tasks.
  3. MIT Project NANDA, The GenAI Divide: State of AI in Business 2025: coverage of the report.
  4. Gartner, press release, July 29, 2024: “at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025”.
  5. Ryseff, De Bruhl and Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RAND, 2024: report.