FOB Cost Predictor
Estimates what a product that does not exist yet will cost, so it can be budgeted and negotiated with suppliers before it is made.
The problem
Estimating a new product's cost was handcrafted: everyone did it by eye, with no common reference to budget, negotiate with suppliers or spot prices outside the reasonable range.
What I did
Designed the predictor's statistical method and its empirical testing: attribute clusters with a custom relevance score (between-group coefficient of variation, R² and Cohen's d), Student's t for small samples, exponential recency decay and CLIP visual cold-start. Tested the method's business hypotheses against the full data history.
The outcome
Empirical analysis over 57,000+ products, presented to C-level. Its infrastructure is now reused by other engines in the ecosystem.
How it estimates
- New product attributesInput
- Comparable attribute clustersAgainst history
- Custom relevance scoreHow much each comparison is worth
- Student's t on small samplesWhen comparables are few
- Recency factorExponential decay
- Estimated FOB costOutput
There is no history for a SKU that doesn't exist yet: the method finds comparables and weighs how much each comparison is worth. When even attributes fall short, cold start is resolved by visual similarity with CLIP.
Stack
- Python
- Streamlit
- SQL Server
- SciPy
- CLIP
- Qdrant
- Supabase
- Docker
Blue Star Group internal project — the code is not public.
Next project: Cash-Audit Fraud Detection →