You might be checking if it’s somehow December or January because that’s when lookback/forward reflection-type blog posts usually get published. Don’t worry. It’s September. But it’s a specific day in September: exactly one year ago from when we launched the foundational pieces for the new Nextmv platform. (It’s also exactly 100 days before the end of the calendar year, but that’s beside the point.)
Since that release, the Nextmv platform has come a long way from a flexible optimization modeling SDK and CLI to include deployment infrastructure, testing tooling, decision apps, and optimization integrations. We wanted to take a beat and reflect on all the goodness that’s been released since then. Take a look, try things out, share your feedback, or send us a message if you’d like to chat.
Testing and experimentation on decision models
Testing decision models help algorithm teams create a clear, repeatable path to production, get stakeholder buy-in and confidence in changes, and avoid expensive mistakes. But testing decision models is hard. This is because there aren’t many out-of-the-box tools that easily integrate with decision model workflows. Often teams are faced with investing in technical setup and maintenance, wrestling with complex model analysis, repeatable presentation of results to stakeholders, and ever-present problem drift.
(Ryan presented on this topic last week with Nicole, our VP of Engineering at DecisionCAMP. It’s worth a watch.)
Over the last year, we’ve shipped curated experiences for exploratory batch experiments, acceptance tests for go/no-go decisions, and shadow tests for model performance under production conditions. And it’s cool to see them all work together in concert, especially in the demo below from one of our decision scientists, Dirk.
The foundational pieces that made this possible are powering what’s on the horizon for curated experiences for scenario testing, switchback testing, and benchmarking. Stay tuned for more here!
Google OR-Tools and HiGHS integrations
There is a lot of optimization technology out there. This Ask HN post from Ryan definitely highlights this. We’re on a mission to help algorithm developers — whether they identify as an operations researcher, data scientist, machine learning expert, or software developer — accelerate their processes for building decision models with our DecisionOps tooling for deployment, model management, and testing. That means we’re investing in integrating with more of the tools people know and love for building these models.
So far this year, we’ve made it possible to go from zero to a unique API endpoint for HiGHS and OR-Tools. We’re fans of these popular open source projects and we tip our caps to the project creators/maintainers: Laurent Perron of Google OR-Tools, and Julian Hall and Ivet Galabova from the HiGHS team. We appreciate the work they do in maintaining the code and the communities that go along with them.
Check out these two demo videos for each project.
A marketplace for out-of-the-box decision apps
The Nextmv decision app marketplace materialized with the delivery of the Nextmv Routing app. This app provides a powerful out-of-the-box routing experience for solving robust vehicle routing problems quickly. And when you’re ready to customize — special constraints, unique value functions, add map and traffic data — you can. It’s all powered by our composable and customizable Nextroute VRP engine. The video below gives a brief tour of the marketplace app, but this video gives a more detailed walkthrough showing experiments and customization.
We’ve also recently shipped another app for shift scheduling (more content to come on that next week, but you can preview the docs here). And this is just the start. There are more Nextmv-provided apps on the way and we’re exploring ways to open up this space to more decision app builders!
Deployment, data partitioning, model management
We started 2023 by announcing a new decision model deployment workflow that allowed for users to quickly spin up a unique API endpoint for their decision app with version and instance management. This work made it possible to partition your data in a way that aligns with how you develop models and operate your business.
As a bonus to this, we shipped a feature we call Teams, that allows users to share their workspaces with one another. This is just a first step toward making the process of DecisionOps more collaborative.
There's been so much fantastic development since we shipped the Nextmv platform last September — and there’s more to come before we finish out the calendar year. Our roadmap includes support for more testing types, cool docs and demos for how to use Nextmv as part of your CI/CD flow, additional solver integrations, more marketplace apps, and more. We’re headed to INFORMS Annual in Phoenix in October (hope it’s not too hot…) to chat with folks. If you can’t make it or want to get in touch to learn more, send us a note.
Finally, a huge thank you to the Nextmv team, our customers, and our community for all of the feedback, support, and collaboration.
May your solutions be ever improving 🖖