tidy-predict
A local replenishment forecasting tool built on TabPFN and Gemma that cleans messy order data, predicts product sales ranges and generates natural-language replenishment suggestions. It runs entirely on a local CPU to safeguard data privacy.
Timeline (last 90 days)
tidy-predict is a weekly restocking assistant tool that reads messy order export files and outputs cleaned data, a predicted sales range for each regularly sold product for the coming week, and restocking recommendations.
tidy-predict runs entirely locally on a laptop CPU, and sales data never leaves the machine and never passes through any AI API.
In the week of November 28, the model predicted the rabbit night light would drop 58% to 1,249 units, but 4,452 units were actually sold, far exceeding the busy-week ceiling of 2,592 units.
Of the 20 products, 17 had actual sales within the P10–P90 prediction interval; the median total was 13,386 units, and 19,101 units were actually sold.
In the data cleaning stage, of the 1,067,371 rows in the demo file, 70,254 rows were singled out (including duplicates, shipping fees, returns, etc.), and the remaining 997,117 rows were used as sales data, with every excluded row recorded in issues.csv rather than silently deleted.
Across the 12 test weeks, all 12 drafts generated by Gemma passed the number checker, each taking about 36 seconds.
tidy-predict only forecasts regularly sold products (sold in at least 20 of the past 26 weeks), only makes one-week forecasts, and can only model holidays and promotions indirectly through 'which week of the year' and the same week last year.
tidy-predict's pricing plan is 199 rupees per week with the first month free, but the author explicitly states that this is only a plan, that no one has paid yet, and that there is no revenue.
Running the 1-million-row demo file from the command line on a 2019 laptop CPU took 52 seconds and consumed no API tokens.
In addition to TabPFN-2 and Gemma 4 both being Apache 2.0, the complete prediction pipeline has been verified to run in an offline Linux network namespace.
1 more timeline events
All Facts (11)
The project uses the MIT license, includes 28 tests, and the repository is at github.com/xexpertai/tidy-predict.
50%Unverifiedtidy-predict runs entirely locally on a laptop CPU, and sales data never leaves the machine and never passes through any AI API.
50%Unverifiedtidy-predict is a weekly restocking assistant tool that reads messy order export files and outputs cleaned data, a predicted sales range for each regularly sold product for the coming week, and restocking recommendations.
50%UnverifiedOf the 20 products, 17 had actual sales within the P10–P90 prediction interval; the median total was 13,386 units, and 19,101 units were actually sold.
50%UnverifiedIn the data cleaning stage, of the 1,067,371 rows in the demo file, 70,254 rows were singled out (including duplicates, shipping fees, returns, etc.), and the remaining 997,117 rows were used as sales data, with every excluded row recorded in issues.csv rather than silently deleted.
50%UnverifiedAcross the 12 test weeks, all 12 drafts generated by Gemma passed the number checker, each taking about 36 seconds.
50%Unverifiedtidy-predict only forecasts regularly sold products (sold in at least 20 of the past 26 weeks), only makes one-week forecasts, and can only model holidays and promotions indirectly through 'which week of the year' and the same week last year.
50%Unverifiedtidy-predict's pricing plan is 199 rupees per week with the first month free, but the author explicitly states that this is only a plan, that no one has paid yet, and that there is no revenue.
50%UnverifiedRunning the 1-million-row demo file from the command line on a 2019 laptop CPU took 52 seconds and consumed no API tokens.
50%UnverifiedIn addition to TabPFN-2 and Gemma 4 both being Apache 2.0, the complete prediction pipeline has been verified to run in an offline Linux network namespace.
50%UnverifiedIn the week of November 28, the model predicted the rabbit night light would drop 58% to 1,249 units, but 4,452 units were actually sold, far exceeding the busy-week ceiling of 2,592 units.
50%