BUZORA

Intelligence, orchestrated.

Scroll to orchestrate
IWhat it forecasts

Six measures, fourteen days ahead. It reads your own history and predicts what the next fortnight should look like.

IIHow it learns

It learns by asking questions about your data.

A decision tree asks a run of yes-or-no questions — was yesterday busy, is this a Saturday — and ends with a number. Buzora builds eighty small trees in a row for each measure, and every new tree works to reduce the errors the earlier ones left behind.

IIIHow a forecast is made
IVWhat the numbers mean

Most of the improvement came early.

The first training runs produced nearly all of the gain. After about round 111 the score levelled off and stayed there. Each run builds a fresh model, so running it again does not automatically make it better — that depends on whether the new data carries anything useful.

411 training runs
3,972 days of history
6 measures forecast
350 checks, all passing

83.8%

The headline score covers revenue and units sold

These are the two measures the engine is actually built to serve, and the two it does best on. It is a fair number to quote, as long as you know what it counts.

67.4%

The like-for-like score covers the wider group

Every measure that can be scored at all, averaged together. This is the number to use when comparing one training run against another, because it has been measured the same way throughout.

round 408

The jump was a change in what was counted

The headline went from 65.2% to 84.1% between two runs. Nothing about the model changed — the headline simply stopped averaging in two measures the engine already grades as weak. Measured the same way either side, the run reads 67.4%.

80 vs 411

Two different kinds of round, easily confused

80 boosting rounds describe how a single model is built: eighty small trees, one after another. 411 training runs count how many times the whole process has been run against the available history. They are not the same number and they do not compare.

Careful

These percentages are not a share of correct predictions

They come from the report's own accuracy measure, not from counting how many forecasts landed exactly right. Reading them that way overstates what is being claimed.

VWhere each forecast stands

Accurate is not the same as useful.

Alongside accuracy, the engine measures skill: how much better the forecast is than a simple rule. This matters because a forecast can look accurate simply because the business is predictable. If sales barely change from week to week, using last week's sales may already work well — so the model has to beat that to be worth anything.

What the range means

Every forecast comes with a span around it, meant to contain about 80% of real outcomes. A forecast might read: expected revenue $2,000, range $1,600 to $2,400. That means across many comparable forecasts, roughly eight in ten actual results should land inside their own range. It does not promise that any single result will.

79.2 → 82.8

Measured coverage improved

Across the whole run, the share of outcomes landing inside their range rose from 79.2% to 82.8%. Some individual ranges still sit below the 80% target.

Why it helps

A range is what makes a forecast usable

A single number tells you nothing about how wrong it might be. The span is the part you plan against — how much stock to hold, how much cover to roster.

VIWhat this proves

Encouraging in the lab. Not yet proven on a real business.

The record is honest about what it is. Below is what the work established, what was thrown out, and the one thing none of it can tell you yet.

Established

Revenue and unit demand forecast well over fourteen days

These are the strongest results in the run and the ones the engine is built around. Weaker results for the other measures are documented rather than hidden.

Established

Later work went into reliability, not raw score

Removing a feature that was doing harm, making the forecast ranges trustworthy, and using separate models for different forecast days only where testing justified the extra complexity.

Rejected

Several promising ideas did not hold up

They looked good on a single test and then failed to repeat across further runs. Each was dropped rather than shipped on a borderline result, which is how an engine avoids accumulating changes nobody can later justify.

Promising

A shared model helps businesses with little history

For a business with only 90 days of its own data, revenue error fell from 27.71% using its own model to 22.65% using a model pooled across several businesses — about 5.1 percentage points better.

Unresolved

Pooling is a consent question, not a technical one

Learning across accounts means one customer's patterns inform another's forecast. The code works. Whether it may be used is a decision for a person to make, not something to infer from the fact that it improves accuracy.

The limit

All of this comes from synthetic data

These results do not establish that a real customer will see the same accuracy. The next meaningful step is measuring forecasts against actual business outcomes, using the same clearly defined metrics.

Buzora

Coming March 2027

Twenty minutes on a call. We map the systems you run today and show you what Buzora would read across them.

Not ready for a call?

Booked. We'll come back with a couple of times.

Two founders. A shared ambition.

We’re two high school students bringing an entrepreneurial mindset and a passion for technology to Buzora, with a shared goal: build something that makes a difference.

03Hardware

It counts the room without watching it.

The customer tracking module clamps over a door frame and reads the doorway below with 24GHz radar. No camera, no images, nothing recorded that could identify anyone. It gives the forecasting engine the one number a till cannot: how many people actually walked in.

Customer tracking module — Product animation

The components

Sensing

LD2410B 24GHz mmWave

Detects human presence and motion. It reads movement through the air rather than light, so it works in the dark and does not care what anyone looks like.

Privacy

It scans the doorway below

24GHz radar. No camera and no images, so there is no footage to store, secure or leak. What comes out is a count, not a picture.

Mounting

Clamps over the door frame

No drilling, no wiring and no bracket. It goes up without touching the building, which matters when the building is leased.

Build

Four M3 cap screws into brass inserts

Heat-set inserts rather than threads cut into plastic, so the housing can be opened and closed repeatedly without stripping.

Power

500mAh LiPo, a full day on a charge

The radar mounts facing the floor so the whole doorway sits in view, and the battery carries a trading day without a cable running to it.

Where the first rooms are.

Click a state to enlarge it, drag to pan, and scroll to zoom.

UNITED STATES
Business locations
Drag to explore

Coming soon.