Most analytics questions look like they need a query and actually need an identification strategy.

I build things that make the difference measurable: synthetic data with planted ground truth, so estimates and agents can be scored rather than argued about.

See the case studies Read the About page

Track record

Bell 2023–2025 · Meta 2025–present

$2M

cost savings from replacing third-party crowdsourced data with an internal solution

28%

reduction in congestion-related network impact

100+ hrs

of manual reporting eliminated per event cycle through the FIFA World Cup reporting pipeline

40+

users of Bell's Data-as-a-Service dashboard

Case studies

Source · ecommerce-promo-causal / results/estimates.csv

01 · Causal inference

A promo that looked like $23 per customer made $6

A naive difference in means overstated a free-shipping promotion's effect by 3.6×. Cross-fitted AIPW recovered the planted truth to within 9%, and the economics flipped the recommendation.

Naive

$23

AIPW

Error vs truth

Read the case study →Code Live demo

Source · ecommerce-analytics-agent / results/comparison.csv

02 · AI evaluation

Does agentic complexity actually beat a fixed workflow?

A benchmark that grades an analytics agent's conclusions against known causal truth: not whether the SQL ran, but whether the answer was right and whether it overclaimed.

Read the case study →Code Live demo

The arc

Studies

What is likely to happen?

Bell

Why is it happening?