Forecasting AI and usage-based spend
Usage-based AI tools do not behave like seat licenses, and most budget forecasts assume they do. I built ai-spend-intelligence to answer three practical questions from a raw corporate card feed: which vendor a charge belongs to, whether that vendor bills per seat or by usage, and how early in the month spend can be forecast with any confidence. The short answer to the last question is around day 12. Before that, a month-end projection on metered spend is mostly noise, and the useful output is a calibrated range rather than a single number. My forecasting background comes from revenue forecasting at Geidea, Saudi Arabia's largest payments processor, where the forecasting work informed a tiered merchant retention strategy.
What I've done
- Mapped 231 raw card descriptors to 15 vendors with 100% purity against ground truth, using a registry match followed by trigram clustering.
- Recovered five AI vendors (Anthropic, Cursor, Perplexity, Replit and ElevenLabs) through clustering alone, with no registry entry for any of them.
- Identified seat, usage or hybrid billing for 14 of 15 vendors from three time-series features.
- Showed that plain run-rate beat a weekday-pacing model with a growth trend, which did worse at three of five cutoffs; at day 5 the 80% interval was plus or minus 100%.
- At Geidea, forecast revenue and modeled merchant churn (ARIMA, logistic regression) to inform retention strategy on a $5M+ revenue portfolio.
Public work
- AI spend intelligence from a raw card feed: Metered spend cannot be called before roughly day 12; on day 5 a month-end projection carries a plus or minus 100% range.
- Queue economics for a fraud and compliance review team: Treats review headcount as a budget line: the marginal value of capacity drops from $2,882 a month for a fifth hire to $109 for a sixth.
- Marketplace payout reconciliation: Compares payout-timing policies by cost: holding the riskiest 10% of sellers for 30 days cut unrecoverable loss to $7,604, against $15,833 for a blanket two-week hold.
Contact
Email sabasharat.ali@gmail.com, or find me on LinkedIn and GitHub. Full history on the experience page.
Open to roles in New York or remote US.