Not too long ago, my cofounder Ryan posed a question to Hacker News: “Do you use an optimization solver? Which one? Do you like it?” The response and engagement took us by surprise. The post even made the top 10 for a hot minute.
What we took away from that experience was: 1) there is a large ecosystem of optimization technology, 2) there is an active and growing decision science community, and 3) everyone has their own preference for how to solve optimization problems. We held these learnings close as we built out Nextmv in 2023 to help accelerate how decision science teams test and deploy many types of custom decision models, and keep them in mind as we look to execute on our planned 2024 roadmap.
While Nextmv has been likened to many things — from the AWS of operations research to the Hugging Face of decision science to the MLOps platform for decision algorithms (DecisionOps!) — we’re focused on building a product that meaningfully accelerates decision science teams and the broader community at large. In the spirit of previous new year posts, let’s dive into the rundown of what got us here and what we’re roadmapping to get us there in 2024.
2023 lookback at Nextmv
When I look at all our team, customers, and community accomplished in 2023, I’m speechless. I’ve been fortunate to reflect on this a few times recently, notably with the one-year anniversary of the Nextmv platform launch and our 1.0 release. In light of those posts, I’ll keep the recap below on the shorter side.
Decision science testing framework and tooling
We shipped a complete suite of testing tools that tie neatly into a cohesive testing framework designed for decision science models. (With all the box and whisker plots you could want!) The suite includes historical tests — batch, scenario, and acceptance tests — and online experiments — shadow and switchback tests (which has had a lot of interest in the community!).
Testing tools ensure quality code and de-risk model rollout to operational environments. Unless you’ve built these tools yourself, these capabilities are a challenging gap for most teams to cross. We’re really pleased to make our suite of composable testing tools available through our platform.
One of the coolest demos (I think) shows both an acceptance test and a shadow test (and ties into CI/CD — more on that below).
Google OR-Tools, Pyomo, HiGHS integrations
We shipped three OSS optimization integrations in 2023: HiGHS, Google OR-Tools, and Pyomo — which slipped in right at the end of the year. We’re excited to demo the Nextmv Pyomo integration on a live techtalk this month.
We also made strides on many other open source and commercial solver integrations throughout the year that we look forward to unveiling soon. If there’s an integration we should consider next, let us know!
Decision model deployment, versioning, CI/CD, and collaboration
I’m lumping together several really amazing feature developments under the banner of operationalizing decision models. In 2023, we made it possible to stand up a set of custom API endpoints for a decision model, manage instances of models the way you want (by geographic region, environment, etc.), set up CI/CD workflows, and share experiment results and access to decision apps for better collaboration and understanding across stakeholders.
Again, to all the teammates, customers, community members, partners, and beyond: thank you for realizing so much goodness in 2023. We’re truly grateful 🙏.
2024 preview for Nextmv and decision science
Where 2023 made strides on deployment and testing capabilities, 2024 is focused on capabilities for operating and monitoring decision science models — from more integrations with other platforms to more visibility and control across your workspace and team. While our roadmap may change as we get feedback from the community, here is a preview of how we’re thinking about the year ahead of us.
More integrations for modeling, solvers, SSO, and data warehouses
We’re on a mission to make all of the awesome optimization tooling available today (and tomorrow) work better and harder for decision science teams within their preferred tech stacks. Grab the modeling framework you want, the solver you want, hook the inputs and outputs up to the data warehouse you choose, et voilà, magic. This means expanding our universe of integrations.
Open source and commercial modeling/solving integrations: PuLP and beyond!
Several open source and commercial solver integrations are on the horizon. We plan to continue exploring opportunities with Python-based tools such as PuLP and NumPy, providing support for Java-based decision algorithm development, and furthering our efforts with commercial solver integrations.
Enabling SSO via Okta, etc.
As more decision science teams explore the Nextmv platform (especially from larger companies) we know the security and ease-of-use that SSO provides is a useful value-add. There are a lot of options available on the market: Auth0, Duo, Okta, etc. We’ve spoken to several Okta users, but want to hear from more folks about what SSO integrations would be helpful. Let us know!
Data warehouse ecosystem: Snowflake, Databricks, Dataiku, etc.
There’s a common data flow for any decision service: inputs for platform data, operational data, predictions, and business rules feed into a decision app that generates plans or solutions and output to operational tooling and subsequent customer apps. We see an opportunity to streamline these service interactions with Nextmv by building dedicated integrations with data warehouses and monitoring solutions such as Snowflake, Databricks, and Datadog. These more direct layers can simplify the orchestration of pushing data to APIs and speeds up the front-end aggregation and back-end storage that are involved with any decision service.
Bring your own visualizations
We love a good visualization. Pictures are powerful catalysts for shared understanding of what's going on in a model. There’s also great value in visualizing data the way you want. Customization is key.
We’re investing in a space for sharing and discussing more decision science visualization — some functional and some just for kicks and curiosity. (Check out this 3D routing viz and some experimentation from Marius below with 3D packing visualizations.)
Across our customers and the broader community, there’s fantastic visualization work happening all around us. We’re working on ways to unlock the ability for anyone to bring custom visualization to the Nextmv UI — starting with a soon-to-be-announced release this month!
Expanded deployment options, richer DecisionOps experience
A user can go from zero to a set of custom API endpoints for their decision model in a matter of minutes. It’s fast. It just works. Whether you’re running in the cloud or self hosting, we want Nextmv to be where you need it. We’re investing in ways for customers to choose the compute to run on and configuration to run with.
From a DecisionOps perspective, we’re working on expanded CI/CD workflow support (via offerings such as a pre-built GitHub Action) and webhook support (for simplifying integrations with tools like Slack to make communicating updates like experiment results easier with Nextmv).
There’s a whole ecosystem of fantastic tools to consider and would love to explore. We welcome your feedback!
Enhancing vehicle route optimization with Nextroute
Many of us at Nextmv got our start in the vehicle routing space, which led to the development of our own VRP engine that we call Nextroute. What we like about it: how easy it is to get started with common constraints and quickly add in map data (because that can be its own process 😬), it’s customizable to your constraints and value function, and it seamlessly integrates with all of Nextmv’s deployment and testing capabilities.
We shipped a steady stream of performance improvements to the Nextroute engine in 2023 and have more planned for 2024. Looking forward, we’re working on making Nextroute more accessible to Python developers, continuing to improve its performance, and delivering on even more visualization goodness I mentioned above.
Unifying, amplifying the decision science community
Whether you know the practice of automating the search for more efficient plans as operations research, or decision science, data science, mathematical optimization, or some other term, it is an incredibly broad, mature, and important layer that underpins global industry.
But despite the prevalence and history of decision science work, the community of industry practitioners still seems to be coalescing. (It’s been especially interesting to see the positive reaction to our efforts to aggregate job postings in the industry!) At Nextmv, we want to expose, surface, and amplify all of the amazing people and work that is happening in this space, while also simplifying how developers access and use the ecosystem of technologies available. We hope this roadmap preview helps give you insight into how we hope to contribute to the community — and we welcome your feedback along the way.
Here’s to 2024! May your solutions be ever improving 🖖
