FROM WORKFLOW TO WORKING TOOL

Data, analytics & ML evaluation

Clean data first, then see it, then check whether your models are really as good as they look.

Discuss your project What you receive

Written scope. Tests on your material. Source code delivered.

Select a step to explore the workflow.

The problem

Exports and spreadsheets come in different formats, with duplicates and missing values. Dashboards show charts before anyone has checked the data underneath. A new model looks better than the old one because of one lucky train/test split.

What you get

Data cleanup: messy exports turned into consistent Excel, CSV or JSON, with the checks written down so you can repeat them
Dashboards built with Streamlit, or a scheduled report in Excel or PDF
ML evaluation with repeated cross-validation instead of a single split
Bayesian model comparison with a region of practical equivalence: is a difference large enough to matter?
Calibration checks and explanations (SHAP or LIME) of which inputs drove a prediction

Deliverables

A report with the method, results and limits
Reproducible Python code with fixed seeds and configuration
The data checks, written down and runnable again

How we measure it

Data checks are counted line by line before any chart; key figures are compared with a separate reference calculation, as in our KPI dashboard demo. Model comparisons use repeated cross-validation and report intervals, not a single score.

Limits

Work at pandas and Excel scale; no data warehouses or BI (business intelligence) platforms.
Evaluation of medical or other regulated models is statistical only, never a clinical or regulatory validation.
No student theses or coursework.

Related work

Synthetic data

KPI dashboard that checks your data first

Sales and expense CSVs checked line by line before any chart, then revenue, margin, cash and overdue invoices.

Test results and limits

171 of 171 planted faults found, 0 false alarms, KPIs equal to a separate reference calculation to the cent on 4 of 4 synthetic sets (measured 5 October 2026).

Read the project

Synthetic data

Bank reconciliation that explains itself

Bank-statement lines matched to invoices, each with a confidence and a reason; unclear cases go to a review list.

Test results and limits

0 wrong automatic matches on all 8 synthetic sets; on the harder stress sets about half the payments went to review instead of being matched (measured 5 October 2026).

Read the project

Insight

One lucky split: how to tell if your model is really better

A single train/test split can make a weak model look good. Repeated cross-validation and a Bayesian comparison give a more careful answer.

Bring your workflow into focus.

Send up to 3 samples or describe one process. We reply by email with what is feasible and how we would measure it.

Start a project