Case Studies
Do you want some more concrete examples of what Mirabolic has accomplished in the past? Say no more. Here we give you a taste of work in three distinct application areas. We love nothing more than getting to work on something new to us.
Creating a predictor for the sale price of homes.
A real estate tech startup needed to predict future home sales prices across the U.S. With questions such as "Why did this house sell for $10k over the asking price?", they also needed help defining what that might mean both in terms of a business product and in terms of what was feasible given available data. With technical goals in need of clarification and even project feasibility in question, we got to work.
-
01. Assess and Build
A sales price predictor (aka an AVM) was developed based on past price data. (If you've seen Zillow's Zestimate, it's the same kind of thing.)
-
02. Evaluating the Performance
We quantified the accuracy of the model. For future sales, our model's accuracy was
essentially identical to the best commercial state-of-the-art pricing models used by
banks. -
03. Getting to "Why?"
We helped the client pin down their definition of explainability, and we created an explainer that automatically determined which aspects of a home were most significant in its price.
-
04. Cloud Architecture
The cloud architecture was constructed by building, storing, and loading models efficiently and reliably.
-
05. Improving Models
We built a highly parallel, distributed, cost-efficient hyperparameter searcher to run automated experiments at scale. It significantly improved the model's accuracy.
Detecting influence and disinformation campaigns in social media platforms.
We have joined DARPA’s project Influence Campaign Awareness and Sensemaking (INCAS), supporting the evaluation effort.
The project performers are designing software to detect influence and disinformation campaigns in social media platforms. Our work involves designing metrics to determine how good a job the performers are doing at detecting the campaigns. Ground truth is both difficult to get or even define, so evaluation is challenging.
Moreover, metrics need support from about 10 separate stakeholder organizations, some of whom have competing interests. Facilitating the conversations to arrive at a standard approach presented its own challenges.
Streamlining Sickle Cell Disease treatment approval with expert statistical analysis.
A large pharmaceutical company had developed a novel treatment for sickle cell disease (SCD), which affects about 100,000 Americans annually.
The company was applying for FDA approval for a Phase 3 clinical trial and needed expert help with the statistical aspects of the experimental design. Because of the costs associated with such a trial, in addition to the associated delay in time to market, the company was interested in determining the fastest possible test design that could achieve sufficient statistical power to convince the FDA. Mirabolic was able to leverage some sequential test design ideas developed in WWII, largely underutilized since, and adapt them to cut the length of the clinical trial in half.