← Back to work

AI · Case study · Private

ParticleX

A synthetic experimentation engine - run A/B tests on AI-generated user personas before risking real traffic, real money, or real weeks.

LovableLLM-based simulationPersona modeling
internal
ParticleX preview

Context

The setup before I touched it.

Real A/B tests are a luxury. They need traffic volume, engineering effort, and weeks of waiting for statistical significance. So in practice, only the biggest decisions ever get tested - the headline of the homepage, the price of the main plan. The other 95% of decisions (button copy, onboarding order, which benefit to lead with) get made on instinct, in a meeting, by whoever spoke loudest.

How it came together

Step by step - expand any phase for the highlights.

  • Mapped the cost of a real A/B test
  • Listed decisions teams skip testing
  • Identified small-team blind spot
Frame the gap preview

Problem → Approach → Result

The short version, for the broad audience.

Problem

  • 01Most product and marketing decisions never get tested - the cost of a real experiment is too high relative to the stakes of a single line of copy.
  • 02Even when teams do run experiments, they wait weeks for results and only learn 'A beat B' - never *why*, and never which segment cared.
  • 03Small teams without traffic (early-stage startups, internal tools, B2B with 200 users) can't run statistical tests at all - they're flying blind by definition.
  • 04Stakeholder debates collapse into opinion vs. opinion because there's no cheap way to get evidence in the room.

Approach

  • 01Generate a population of synthetic users with realistic demographics, intent, motivations, and decision biases - not just demographics, but the *why*.
  • 02Run any variant (a piece of copy, a landing page, an offer, a pricing tier) past that population and aggregate predicted behavior at scale.
  • 03Surface uplift estimates *and* segment-level reactions - so you don't just see 'B won', you see 'B won because price-sensitive users finally understood the discount'.
  • 04Position the output as a pre-flight check, not a replacement for real testing - point teams at the 2-3 variants worth burning real traffic on.

Result

  • 01Cycle time from idea to directional answer dropped from weeks to minutes.
  • 02Low-stakes decisions that used to be argued about now get a fast evidence-shaped answer and move on.
  • 03Teams enter real experiments with a sharper hypothesis and fewer variants - so the experiments they do run are more likely to produce a clear winner.
  • 04Smaller teams without statistical traffic finally have *a* signal where they previously had none.

Role

Concept, design, and build - solo from zero to working product.

Status

Live - internal

Next

Browse more work →

React

Tap if this resonated.

What do you think?

Leave a comment or ping me — any feedback, thoughts, or collab ideas, I'll really appreciate it. Building alone is no fun.