August 2023. The downstream application of Project G v1 — closing the loop between food reformulation and market reception. Where v1 reverse engineered ingredient quantities from nutritional labels, v2 asked the harder question: will consumers actually like the reformulated product? Built, then cancelled by the client.
context
Project G v1 gave the client the capability to reformulate food products from first principles. But reformulation without validation is half the problem. Knowing what you've made doesn't tell you whether anyone wants it.
Traditional methods to answer that question — focus groups, sensory panels, post-launch data collection — are slow, expensive, and arrive too late to de-risk a launch. The LLM wave in mid-2023 created renewed appetite for using AI to shortcut this process.
what we built
A machine learning model to predict coffee favorability based on consumer demographics and coffee characteristics — a prototype for a broader consumer-reception prediction system internally codenamed Regenesys v2.
The data pipeline:
- Make coffee to defined characteristics
- Have consumers taste it
- Have consumers rate it
- Collect consumer demographic data
- Collect coffee characteristic data
- Combine into a training dataset
- Train a model to predict favorability scores
The model treats the problem as a supervised prediction task: given a consumer profile and a coffee's characteristics, what rating would that consumer give?
the idea
The goal was to de-risk new product launches. Rather than running focus groups or waiting for post-launch signals, a trained model could predict reception for any new formulation before it was manufactured at scale — giving the client confidence to reformulate boldly.
Combined with v1's constraint-based reformulation engine, the full system would form a closed loop: reformulate → predict reception → reformulate again → repeat until the predicted favorability hits target.
what happened
Built it. The project was cancelled by the client before it could be deployed.