Know what comes next, not only what happened

Data Science & Applied ML. A report tells you what already took place. A well-framed model anticipates demand, ranks priorities or checks what nobody has time to look at. I build predictive and analytical solutions for e-commerce and logistics, and computer vision projects, measured against the business outcome.

When it makes sense

Buying or production runs on gut feeling, and inventory pays for it.
Years of stored data that nobody uses to decide anything.
A repetitive visual check relies on nobody getting distracted.

What I deliver

Demand and stock forecasts
Predictive models for e-commerce and logistics
Computer vision for detection, counting and inspection
Model in production with performance tracking
Results report in business language

What I work with

Python
SQL
Machine learning
Computer vision

What it looks like

Examples built on synthetic data to show the shape of each deliverable. None belongs to a client.

Demand forecast

Twelve weeks of history and six projected, with what drives the number and where the risk is.

Synthetic data

Service level: 97.2 %; +4.5 pp vs. previous method
Forecast error (MAPE): 8.4 %; −3.1 pp vs. previous method
Stockouts avoided: 38; per quarter
Idle stock: −14 %; vs. last year

Weekly demand, in units

What drives demand

Seasonality 34 %
Promotions 27 %
Price 18 %
Weather 12 %
Holidays 9 %

Stockout risk, next six weeks

W 13 W 14 W 15 W 16 W 17 W 18 Beverages Snacks Dairy Cleaning Home Pets

View the data as a table

Weekly demand, in units
Week	units	80% band
1	820
2	845
3	810
4	870
5	905
6	880
7	930
8	960
9	940
10	985
11	1,010
12	1,040
13	1,065	± 40
14	1,090	± 55
15	1,075	± 70
16	1,120	± 85
17	1,150	± 100
18	1,170	± 115

What drives demand
Factor	Weight
Seasonality	34 %
Promotions	27 %
Price	18 %
Weather	12 %
Holidays	9 %

Stockout risk, next six weeks
Family	W 13	W 14	W 15	W 16	W 17	W 18
Beverages	very low	low	low	medium	high	very high
Snacks	very low	very low	low	low	low	medium
Dairy	low	medium	high	high	medium	low
Cleaning	very low	very low	very low	low	low	low
Home	very low	low	very low	very low	low	medium
Pets	medium	low	low	very low	very low	very low

Computer vision inspection

Counting parcels on a belt and flagging the ones to check.

Synthetic data

Photo of cardboard boxes on a conveyor belt, with simulated detection boxes over each one and one flagged for review.

Real photo, public domain (U.S. Department of Agriculture). Detections simulated for this example.

Counted: 128
To check: 3
Precision: 96.2 %
Recall: 94.8 %

What the model saw and what it was

Actual / Detected UndamagedDentedUnreadable label Undamaged 412 6 2 Dented 9 71 3 Unreadable label 1 2 38

Parcels per hour

View the data as a table

What the model saw and what it was
Actual	Undamaged	Dented	Unreadable label
Undamaged	412	6	2
Dented	9	71	3
Unreadable label	1	2	38

Parcels per hour
Hour	Passed	To check
8:00	96	2
9:00	112	3
10:00	120	1
11:00	118	4
12:00	104	2
13:00	98	3
14:00	122	2
15:00	126	5
16:00	119	3
17:00	110	2

Does your process stall on this line?

Two lines on what runs late are enough to start.

Tell me about it