AI Engineer NYC 2026
Join AI Engineer and Major League Hacking (MLH) on Sunday, October 11, for a full day of building, experimenting, and meeting fellow AI engineers.
Ripple
A Bayesian what-if simulator for pricing, promotions, and new-city launches that updates its forecast as real launch data arrives — for team
About the project
Problem
Teams decide on promotions, prices, and new markets with one number, no uncertainty, and no check after launch.
What Ripple does
- Simulator: answers “what if” with incremental profit, a 90% interval, and the chance it pays.
- Launch tracker: each real week updates the forecast with Bayes’ rule (PSIS reliability gate). Austin: 25% → 94% → 100% chance it pays.
- Ask Ripple: an AI agent that explains results; a number guard blocks invented numbers.
- Your data: upload a CSV and fit the model.
Model
Hierarchical Bayesian demand model (PyMC): price elasticity per group, promotion lift and dip, seasonality, and city traits for new cities.
Proof
6/6 trust checks, 30/30 parameters recovered, calibration 90–94% for a 90% target, 174 tests.
Limits
Synthetic demo data with known truth; the market scan is a labeled sample.
Built with
AI tools
Prize Categories Entered
- Event prizeSubmitted
Best Overall
Ripple answers "what if?" business questions with honest uncertainty, and it learns from real launch data. Technology: a hierarchical Bayesian model (PyMC) gives each answer with a 90% interval and the chance it pays. Each real launch week updates the forecast with Bayes' rule and a PSIS reliability gate. A number guard stops the AI agent from inventing numbers. Proof: 6/6 fit checks, 30/30 parameters recovered, and 90–94% calibration for a 90% target. Design: clean UI; sliders update instantly (Austin launch: 25% → 100% chance it pays). Completion: live at ripple-opal.vercel.app, 174 tests in CI.