Anytime you open up Google Maps and, like, the transit option takes an hour and, a half, but you can drive there in 20 minutes, there’s a question of, like, “Hey, could there be a better way?” Stop Requested.
This is Stop Requested. by ETA. I’m Christian. And I’m Levi. These are real conversations with the innovators, operators, and advocates driving improvements in public transportation.
Today, we’re talking with Anthony Tresati, co-founder and CEO of Antra Mobility, about how transit agencies can use demand modeling, optimization, and AI to make better service decisions. Antra’s technology allows planners to test thousands of potential network designs, compare trade-offs between ridership, coverage, cost, and equity, and identify practical changes that could be implemented over time. We’ discuss how Anthony’s work on
Marta Reach developed into Antra, why transit planning requires more than simply generating a technically optimal network, and how agencies can use AI without turning critical decisions. over to a black box.
Here’s our conversation with Anthony Tresati. Welcome back to Stop Requested. Today, we have a very special guest that’s joining us, the CEO and co-founder of Antra’ Mobility,
Anthony Tresati. Anthony, how are you today? Yeah. Well, it’s really our pleasure. I– This is going to be a lot of fun. I’ve followed you online.
Uh, I’ve seen what you’ve been up to. I, I know that, you know, you have some connection with YC, and that’s Y Combinator for, for folks who may not know.
Um, your, your background is really impressive. Uh, and I probably should be referring to you as doctor in the intro as well because y-you’ve got your PhD, you’ve done a lot of research.
Uh, but for those who don’t know you, uh, and they’re– maybe they’re just finding out about Antra for the first time, could you give us a, a bio on yourself and on Antra? Yeah.
Yeah. So my work in public transit started when I was at Georgia Tech, where I was doing my PhD. That’s where I met my co-founder, Connor, as well as our, our designer,
Sam. We were all working together on a, a number of projects there, a lot of them with Marta in Atlanta. So I was doing my PhD in operations research focused on public transit optimization. A lot of that focus was on data-driven network design of transit systems.
So looking at how people are using the current system, how people are traveling not using the current system, and where there’s the opportunities to get more people using transit.
Um, so Antra is an AI platform for transit planning. Uh, we, we started it, we, after we left Georgia Tech, me and my co-founder actually both went to Google for a little over a year, and then we started Antra after that.
We got funded by YC, so then we quit our jobs and have been full-time on it for the last two years. Uh, and we’re really trying to help agencies answer the questions of, you know, given your resources, what’s the best service to run? How do you maximize ridership, efficiency, all within a fixed budget?
A-and so that’s, that’s really what led you down that path. Um, it sounds like you had that particular question in mind. M-m-maybe it’s because of agencies being, uh, short-staffed or underfunded.
I, I mean, w-we’re– if we dig a little bit deeper into that problem statement that you identified, like what, what drew you, uh, to trying to solve that problem in particular? Yeah, I think one of the biggest things that’s just a very striking point is, like, anytime you open up Google Maps, and, like, the transit option takes an hour and a half, but you can drive there in twenty minutes, there’s a question of, like: Hey, could there be a better way?
Is there enough, is there enough demand that we could add a service to capture more? Or how, how can we do things more efficiently? One of the big things that we were exploring while we were at Georgia Tech was a lot of the on-demand multimodal transit systems.
So in the case of Marta, you have, uh, you know, you have your backbone of the subway that’s running north-south, east-west, and a, a number of bus routes. But there’s a question of, can you add an on-demand service in certain areas to get people to and from the train station, to and from the bus stop, and get people where they wanna go faster?
And so that’s where a lot of the research focused, uh, early on. Uh, my research specifically focused on how do you design the bus network around the fact that you’re going to be using this on-demand service. So we did a lot of this long-term planning of what might the future of Marta look like twenty years in the future if you had transformed more into this on-demand system where you get a ride through the transit system, and then you take these high-frequency, high-occupancy vehicles to and from, uh, along the high-density corridors.
Uh, and Marta was actually pretty interested when we presented that future network to them, and then that actually led to a pilot. Um, uh, uh, all this work was done under my advisor, Pascal van Hentenryck. So he helped get a grant and a collaboration with Marta to actually launch Marta
Reach. And I helped design the rider applications and driver applications, and Connor did all the cloud infrastructure and cloud dispatching. Uh, and we actually operated this Uber-like service for public transit. So for the same two dollar fifty cent fare, you get a ride to or from the train station or the bus stop, plus that transfer to get you where you need to go. So
A-Anthony, let me– First of all, that, that’s remarkable. That’s, that’s all remarkable and, and, uh, it’s great work. I, I love, uh, when we’re doing, uh, what you call data-driven network design. I, I, and I really like that, uh, because in a lot of communities, we find that, you know, sometimes some of the decisions is political of, you know, how the service is designed and, and how the resources are being used because, you know, we wanna put, like, uh, mostly, like, those coverage, networks. Like, oh, just put a little bit of everywhere, and there’s at least that life land, uh, lifeline or, or safety net for people in the community to be able to get to places.
But then the, the service itself is not very useful or, like, ridership-driven, right? Like, it’s not gonna attract a lot of people. And you were saying people check on Google Maps, and you look how long it takes to travel in transit and in driving a car, and it’s like transit sometimes in certain places is, like, three times what it would take driving. So it’s, it’s just, like, frustrating and, and is it
Thinking, is it the right design? Like, are they, are they looking, um, you know, a, a, at their, uh, community, the mobility needs, and then making data-driven network design decisions, right? Of how the service should be established. So going to Marta, and, and I wanna understand a little bit more, uh, the tech stack and then some of the things that, that, uh, you developed there and, and I understand that was while you were going to Georgia Tech, is that correct? Yes. Yeah.
And, and then for… So is it they created one app that is multimodal in multi-planning, including microtransit kinda Uber-like, or you created this Marta
Reach was just an app that is exclusively for like microtransit, that, that Uber-like service you were describing? Could you tell me a little bit more about their, their technology stack and the services, how tho- those two things relate? Yeah. Yeah. So for the pilot, I, I had developed the, the Marta Reach application. So we developed a rider app, we developed a driver application, and then, and then
Connor developed the cloud infrastructure, and then we had a whole team of other people as well. Um, but I, I led a lot of the development on the rider app and driver app side.
And, and for the initial rider app, it was much f- it, it was very focused on the on-demand piece where you, similar to a Lyft or Uber, you just book the trip in the app, and you can book it to one of the nearby train stations or bus stops. And then the car c- and then you see the car in the app and it comes, picks you up, drops you off.
Um, our, our Auntra Mobility app does actually do multimodal trip planning, uh, and, and we have rolled that out since.
Um, but the, the pilot was focused on the rider application, driver application, both built in React Native Expo, uh, as well as having the, the cloud dif- dispatching backend on Azure hosted. So, so, um, uh, Marta
Reach was that on-demand, and you said you guys created the application for the driver and the application for the rider so they could, you know, give, give, I would imagine, turn by turn and all the information to the driver, kinda like- Exactly … Uber and Lyft does, and then one for the riders. And, and did it have the features for the most part, like, you know, real time, uh, making trip requests, but were you able to also book in advance? Was that a capability of it? Uh, this was back in twenty twenty-two, so we didn’t have booking in advance-
Oh, okay … uh, initially. Um, but we, we made a bunch of improvements as, as we went. Um, the, it was all real-time information in terms of, you know, the driver would get, “Hey, this is what your next stop is,” and then if someone else made a request as they were going, then once they made that pickup, then they would say, “Oh, now pick up this next person before you go to the train station.” So we were giving that, uh, kind of stop-by-stop optimization, optimized route to the rider or to the driver, and then also displaying that information for the rider to see when is the expected pickup, all that information.
Could you share a little bit, um, a- about downloads or, like, the, how many people were using it? What, what was the success like from, you know, when it was first launched? Or and, and if you could recall or share some dates, like, oh, we launched on this date and, and it’s been on since then, or if at any time. has transitioned to something new, and then the usage. Yeah.
So we ran this, uh, six-month pilot back in twenty twenty-two from March to September. Um, and basically what we saw is we, we, we operated this in four zones. We actually have a white paper on it as well, so we have a lot more data there for anyone that’s interested in seeing. Um, but basically it was, you know…
We, we saw basically fifty or a hundred percent growth each month. So w- we, we didn’t know- Wow … you know, where the upper bound was necessarily of where the ridership was going to end. But in the last month, I wanna say we had about, you know, five hundred rides in these four s- smaller zones.
Um, but I, I would have to, uh, I might have to double-check that number. Um- So about five hundred rides a month?
Yeah. Yeah. Um, but now- Per zone … they’ve launched it… Now, this year, they actually just finished their redesign and launched in twelve on-demand zones, I believe, with their new Marta, Marta network, and they, they relaunched the service with another provider.
Uh, but a lot of this motivated, uh, what we’re working on at Auntra. Right. And, and, and also for Marta to see how they can, uh, how a, a previous, um, executive called it address the transit deserts, right? For those microtransit zones will be able to put some form of, uh, transit or connection to transit within, uh, those different areas.
So l- let me ask you. You, you did the research. You come from data and analysis and understanding really, um, you know, the makeup of transit and, and using that information to, you know, again, making data-driven network, designs.
Um, where did the academic version of the problem defer most for the operational reality inside a transit agency? You know, from, from the, from the textbook in, you know, what it says is the, the proper way of doing public transit and analyzing data to, to what you see in the reality of a transit agency.
Yeah. So I, I think a lot of our journey while we were working on the network design piece of it was trying to add in more reality as we went. So, like, the initial network design problem started off by assuming bus routes were just, like, two points, like, a s- a start and an end.
Uh, but obviously, bus routes usually have, like, 40 stops. So over time, we’ve definitely been baking in more and more into our, our latest algorithms in terms of being able to model, uh, the schedule of the bus and how the frequency changes throughout the day. Uh, when you’re designing a new route, taking into constraints of, like, what roads can be used, what the traffic conditions are at different times of day. Um, the academic research was very focused on how do you actually optimize this problem at a large scale, um, and now we’re focused on how do you solve the problem that’s the most meaningful to the rider, uh, and to the agency. So one of the other big differences is, you know, our initial network design, we focus on basically how do, how do you redesign the whole network from scratch? But transit agencies aren’t gonna redesign their whole network from scratch. They’re gonna figure out, “Hey, what are the best things that’ we can change?” So our, our– one of our big offerings at Entre is that we optimize how to do an implementation plan of, like, what are the five best things to change this year, and then what are the five things that stack on top of that the following year, and, and being able to account for both how riders are using this current system and optimizing over how riders would use any given system. This episode is brought to you by ETA. For decades, transit agencies have been locked into legacy CAD/AVL systems built for another era. Expensive upgrades, rigid architectures, and software that lives in server rooms instead of the browser.
Our web-based CAD/AVL delivers industrial-grade reliability with cloud-native speed and flexibility. It integrates with what you already have, replaces closed ecosystems, and gives your team a single real-time operational view without starting over. Replace legacy, move forward. The future of CAD/AVL is open. See it at etatransit.com.
Anthony, is, is that prioritization built into Entre now, or is that, was that just part of the Marta project that you’re working on? It’s built into Entre now. Really? So it, it can, it can give you that recommendation as to what steps you should take and, and when. Uh, I’m, I’m curious, how, how did you kind of arrive at, at that, you know, at that nexus? It, I mean, it sounds like you, you had the previous experience at Marta, and that was pretty informative, but, you know, taking that a step further and then building a, you know, building software around that. Would– Are you hearing that from transit agencies? Like, we don’t know, you know, which, uh, you know, route configuration to tackle first or which schedule to change next. Uh, is that the type of feedback you’re getting? Yeah. I think one of the really challenging questions for agencies is often, “Hey, what should we be prioritizing?” Just even at the agency objective. Should we be prioritizing ridership? Should we be prioritizing coverage? Should we be prioritizing capturing new riders? Should we be prioritizing capturing, uh, additional trips from our existing riders? And all, all these different angles of all, all the different ways you could break down these KPIs and really focus in, being able to use generated network designs for being able to quickly generate network designs for different objectives and, and narrow in on, hey, these are a few weighted combinations of ob- objectives that, that really makes it easier to have the discussion about what should the priorities be. And so often when we’re working with these trans agencies, we try and narrow in over, over the, the course of our, our engagement to basically narrow in on, hey, this is what a more ridership-focused scenario would be, or this is what a more cost-efficient-focused scenario would be, or this is what an expansion scenario would be. If you have ten percent more budget, this is how much more you can do, and you can take that to your stakeholders and say, “Hey, this is why we wanna get ten percent more budget.” So what would you say i– is the hardest question for, transit agencies, uh, to try to answer, uh, with some of the traditional planning tools? So, I mean, we, we know. some of the, the names in the space, uh, you know, and some of them are pretty good. I, I’m curious w- why they might not live up to what the transit agency is looking for.
Yeah. So most of the existing transit tools in the space, the– their work really starts once you’ve already determined what routes you wanna run, what stopping patterns you wanna use. You can draw new routes in the, in the existing tools.
You can, um, you can explore different handcrafted scenarios, but what our, model does is actually explore, you know, thousands, millions, even trillions of potential network designs.
Uh, really, we’re generating thousands of routes and exploring all these different combinations and figuring out what best combinations work for your, both your existing ridership and potential ridership that you’re not currently capturing. And the– what makes us really unique is what we’re doing is we’re doing all this optimization over a travel demand model. So we’ll come in and build, build a travel demand model for the agency and calibrate it so that accurately predicts their current ridership in terms of their route level ridership by hour of day, and we’ll break it down into a number of different ways. And what this does is it, a lot of agencies have a data problem where they don’t initially have a good, clean set of data in terms of where their riders are getting on, where they’re coming from. Like, what, what is the walk to get to the bus stop for your, for your current riders? Uh, and that’s all pieces that we’re modeling, and then that also informs us that when we make the new route or we add in an on-demand light to start, how will that change how the rider uses the system?
Uh- By being able to accurately predict how they’re currently using the system and being able to extrapolate on how they would use any given new system, that allows us to make a, uh, really strong recommendations for how to boost ridership or how to improve efficiency for your current network, and, and, and explore different balances of KPIs in terms of your agency priorities and allow you to help set those priorities and, and quickly explore different combinations of priorities.
Well, I, I just gotta say, Anthony, you know, hearing that, super impressed. It, it sounds, uh, very sophisticated, uh, this model. A, a couple of thoughts, uh, come to mind. I, I’m, I’m not sure which one I wanna pull on first because they’re, they’re so interesting here. But, uh, let me just go with, uh, you know, it– You, you clearly have a, a software that you’re selling, but it also sounds like part of Ontra is also this service model.
Um, why, why take that sort of hybrid approach with transit agencies? Did you feel like, uh, maybe in some of your discussions with them that they, they needed that hand-holding, so to speak, uh, to kinda get them to the place where they could use the data? Or, you know, maybe there’s something
I’m just not seeing there. Yeah, I think it’s a combination. One piece is, you know, it’s how the R-R– how the RFPs are structured, and, and, you know, are they structured as buying consulting work or are they structured as buying software?
Um, so a lot of it’s just a lot of this work is currently done by consultants. Um, and so a lot of our work has actually been as sub-consultants, so we’re in, in some cases, selling our software to the consulting agency that’s then providing that more hands-on service. Um, but we’re also working directly with agencies in some case.
We actually, one of the things we were busy with the last couple months is working, uh, on the Transit Tech Lab POCs with– We’re working on two POCs, one with the MTA’s New York City Transit, uh, on some bus network design, as well as working with the Long Island Railroad on some schedule optimization.
Um, so in those cases, we’re trying to build towards those direct agency relationships and, and being able to ha-have them leverage our software directly. I see. And a-a-another follow-up, um, based on what you were describing earlier, uh, is so you’re, you’re able to, you know, with your algorithms, be able to, uh, see where that latent demand is, uh, in, in a way that the, the human brain is just not able to, to comprehend most of the time. Like, you know, Christian and I have sat in the planner seat and drawn routes, you know, on GIS or, you know, one of the, one of the other tools out there. And we thought we were doing a pretty good job, but there’s just no way that we can compete with the computer.
Um, you know, if you’re looking at millions of points or trillions of points, I don’t know how long that would take for a human to do, but it’s probably e-enough for a lifetime. And a computer can do that within, yeah, I would imagine seconds, right?
Yeah. Um, it obviously depends on the size of the problem. In some cases, we can solve in seconds, but when you’re looking at, like, a New York City, uh– when, when we actually, I think I looked at the, the number of– our, our total search. space was actually ten to the forty-second potential network designs that we were considering. And so optimizing over that, we ran it for about twenty-four hours to optimize the, the one of the, one of the sections of the network. Um, but obviously, New York’s one of the largest transit agencies and, or the largest in the US, so, uh, a lot of the other agencies are much quicker to optimize. Would you, describe then what you’re doing as, uh, scenario-based planning? Or is, is there something else to it? Um, just hearing you talk, it, it sounds like that’s a, a good descriptor. But I, I’m curious what, you know, in your own words, how you would describe Ontra. Yeah, I think that would be pr- an accurate way to describe it. Uh, we’re, we’re creating a, a, a scenario in terms of both a, a calibrated demand model as well as y- what your objectives are, but making it so that it’s pretty reusable, and you can switch your objectives and then explore a new option and get a new best solution, and, and then s-see what the trade-offs are, and, you know, a-and leverage that information in terms of making your decision.
And, and one of the things that also makes us unique is then once you see a solution, you can then react to it and say, “Hey, I wouldn’t actually put a bus route here for XYZ reason.” And then you can actually, you know, modify which stops the optimizer’s considering. You can, uh, lock a particular route. If it removed a route that you think politically is too sensitive to change, then you can just lock that route, and the optimizer actually optimizes around that. So in– if you’re optimizing the buses, you lock all the subways, and the optimizer won’t change any of those schedules. And, and so given that flexibility of, hey, you can focus in on a single route, or you can focus in on an area, or you can fo-focus on the whole network, that allows you to really explore different potential optimization and, and workflow questions. That’s impressive. It, it– but it, it’s, it’s very complex, right? It, it– because this is what I was thinking, um, with a model like yours and, and that it has that intelligence to think about all the different options based on the current demand and projected demand. But there’s so many variables, right? And, and I, I do– as you were saying, maybe you can lock the mo- certain modes and not touch them and optimize outside of it. Uh, but can, can it do pretty much based on, uh, any, um, parameters?
Meaning, like, let’s say I have a given service area, and then I wanna see what would be the best to serve the area without, uh, contemplating any existing service, right? Like if you had just an area and, you know, you wanted to see what would be the best combination of, uh, you know, mobility services for, you know, for a- Yeah, exactly.
Yeah. You can either use the existing routes or you can even, you know, disable, say, “Don’t consider the existing routes. Just- Generate new routes from scratch that are, you know, optimized for the ridership that we’re seeing and, and seeing potential for.
Y-you could even be as specific as, “Hey, we want a route that starts at this point, and then, and goes in this direction,” and then have it explore 100 different options that start at that point and go in different directions. So it gives the planner a lot of control to quickly explore options. Right. Because the, the thing is you also have, uh, infrastructure in place. So you, you- Yeah. … can look at the, at the, the scenarios, you know, uh, without limitations and parameters and just, like, purely based on origin, destination, and projection for mobility demand, uh, give me, you know, what it would look like. I would imagine, and, and I’ve seen in exercises that’ are more manual, that you end up, uh, finding out that a lot of the core portions of your service will remain even if you were to give the
AI, uh, you know, like, no network, create a network because, you know, different, um, factors, including, like, the origin and destination. So a lot of the main networks are close to a lot of the places where people wanna go. So, you know, you, if, even if you were to erase that parameter of, like, or, respect this route or we have these things in place, the model might still put service and frequent service, uh, along the main places where you run today.
Uh, but I think it’s a great exercise, right? To, to show the community the different options and make more data-driven decisions. Um, I wanna ask about optimization and then the use of AI, right? Because, you know, as, as a transit planner, a, you are constantly on that effort optimizing. Optimization is your job. If there, was nothing to optimize, there was no planners whatsoever. Everything would stay the same. But you’re constantly looking at how can we change it, how can we tweak it, how can we prove it, how can we decrease it, or, you know, do whatever is needed to be able to, um, you know, i-improve the service and, and get efficiencies out of it, uh, out of it.
So could you tell me a little bit of w-well, what you know or you’ve seen or any implementations of AI, uh, in the context of transit planning?
Yeah. Yeah. Obviously, uh, there’s obviously, like, the two… There’s many different branches of AI. Uh, there’s…
I’ll break it down. First, the non-LLM cases and then the, the LLM cases. Mm-hmm. Um, so for the non-LLM cases, obviously, like, for just travel demand modeling, like, using, uh, uh, like, a multi-logic model or o-other types of ML models is usually the standard.
You can obviously use neural nets or something as well for basically predicting, you know, when people are going to take transit versus not, and also what transit option they’re going to take. For, you know, in New York, for example, there might be multiple different subway and bus options for getting between two points.
Um, so being able to predict, you know, what route they’re going to take and when they’re going to take transit at all is, is a key focus area, as well as doing the optimization, so doing the mathematical modeling. That’s another form of
AI. Um, in terms of, like, the LLMs, like, being able to use a ChatGPT or a, a Claude, I think one of the things that obviously it lacks in terms of being able to do transit planning is obviously you can’t use something if it’s going to hallucinate. So being able to-
Right … uh, leverage LLMs, I think the most critical part there is having tools that are able to do the demand modeling, able to do the optimization, able to do the simulation, and actually you’re, you’re explaining the simulated results with a verified model as opposed to just relying on, hey, what number did it come up with when it can’t even add numbers correctly. Right. So it’s pretty much trained for transit, right? Like, that, that it knows exactly, you know, what transit is and, and it’s not hallucinating, like you said, is… And, and- Yeah …
that’s the thing, uh, you know, I, I… for planners is, is a lot of times they can, uh, focus on, on only a few models at the time, you know, a few scenarios. So, so how can this modeling and optimization.
he-help planners with more models and scenarios? Yeah. I, I think one of the key things is that you can calibrate the, the demand model and calibrate the, the current system and then lock that piece of like, hey, these are the parameters that we’re using to model our system, and then you can use the AI to then calibrate scenarios or, or configure scenarios of, hey, we want test X, Y, and Z. And then the scena- the agent can maybe have a better understanding of what’s happening under the hood and actually configure the scenario so that you’re testing what you want to test. Like, how do we increase ridership? How do we increase… How do we, uh, find areas to save costs so we can reinvest in another area?
Um, being able to use AI for the explainability on the back end as well as the c- uh, the setup on the front end, I think is, is the place that I see LLMs being the right fit for, you know, the near future at least. And, and, and when it comes, we talk about the predictions and the predictions models and, and the projections, right? Like, for, for, uh, demand, you know, origin, destination, m- you know, um, preference in terms of m-mode of travel and, and so on.
Uh, where do you think agencies need to be more cautious when, when they’re relying on this automatic recommendation? Especially, you know, of course a person is gonna think and, and maybe, you know, come up with a few scenarios based on, you know, what they heard, what the drivers are saying, what people in the community are saying, and they’re gonna evaluate some of the scenarios. When you have the ability to evaluate several scenarios and play with the parameters and the projections, then you get to look more holistically at everything there’ is. But I would imagine When you’re looking at, uh, you’re pushing the parameters, like let’s say the, the projections for ridership and, and for service demand, and then you start maybe tilting, uh, to- towards something that is maybe less probable, uh, then less, there’s less degree of confidence,
I, I would imagine, of a recommendation with parameters that are less probable. So how, how do you play with that? How do you manage that? Where– Is that a place where agents should be cautious?
Like, if they’re, uh, kinda like pushing the parameters to like the end of the spectrum versus like the most probable scenario, is that when they should be more cautious or, uh, what, what, what do you have to say about that?
Um, I think for the most part, it’s not them calibrating the, that piece of it. It’s calibrating, uh, what their objectives are, so what the balance is for their ridership, and coverage, and equity, and reliability, and cost goals are, like those different pieces. And obviously, like, uh,
I, I think you don’t wanna optimize just for one goal ’cause then you’ll, you won’t end up with a network that works for everyone. Uh, but there, there’s a lot of opportunity there to basically when you’re getting this feedback from people, I think one of the things that’s currently lacking, uh, is that there’s not a good way to explore every suggestion that someone comes up with of, “Hey, we, should add a stop here. Hey, we should add a route here. Hey, this route should be higher frequency.”
If you can actually then leverage the AI to triage these issues and say like, “Hey, KPIs look good for this,” or, “This is why the KPIs don’t support this potential idea,” then you’re able to have more discussion with, with the constituents or stakeholders about, you know, what, what things might work and what things, and why some things might not work.
Uh, so Anthony, you know, in, uh, our experience, Christian and my experience, uh, where we have to bridge the gap between planning and, and day-to-day operation, uh, sometimes things just get lost in translation or communications break down.
Uh, we’ve even seen it, uh, where we plan a route, and on paper it’s perfect, right? It’s- Yeah … uh, like e- everything just lines up so well, end of line, and all those transfers between the, the vehicles are, it’s gonna go so smoothly. We got plenty of layover time there.
And, you know, when rubber meets the road, uh, sometimes that. doesn’t turn out, uh, the, the way that you had planned in, you know, either writing it as a, as a test run or behind the desk when you’re drawing the route. I’m, I’m curious if Antra has a, an ability or if maybe you’re thinking, you know, longer term, w- what, uh, what or how are you able to kind of, facilitate the communication between those two departments? Because it really is the, the key, I think, to the transit agency’s, uh, service delivery. Yeah. Yeah, that’s, a great question.
Um, a, a, a few aspects there. One is that’ a lot of the times we’re- modeling these, uh, these wait times as probabilistic, so we’re not assuming that it’s gonna run exactly as scheduled, uh, which helps us create a more robust schedule overall.
Um, but when we’re, when we’re designing new routes and, and designing new schedules, uh, definitely talking with operations is critical, and, and we often loop them in. Um, for, you know, one of our recent train scheduling projects, we had the, the, you know, this, the, the head planner there for the operations side on the calls and, and, and telling us, “Hey, this is, this is how long it takes to turn the train, and this is how long…” Like, th- those pieces of, you know, the critical know-how of what, needs to be considered when you’re doing the planning. Um, but that’s obviously one of those things that, you, know, giving them a full-on schedule and, letting them react to it is sometimes the best way? to uncover, like, “Hey, this is why we don’t think this will work,” and then that allows you to then incorporate that back into the optimization.
So one of the really important things about our model is it’s super flexible and able to incorporate constraints both on like how, you can operate it, but also, you know, how riders are going to use it. So the planning and the scenario-based planning, more specifically, this, this is one part of, you know, how a transit agency puts their service out on the road, but there are so many ancillary technologies that y- you need to consider.
I, I’m not sure y- you know, how far along you are in, in thinking, um, you know, about how you’re, how you’re connecting to perhaps dispatch or, you know, a CAD/AVL system, like, like ETAs or, uh, rider communication and, uh, you know, GTFS that you’re putting off as exhaust from creating the, the, you know, the routes and the trips. Describe to us and our listeners where y- you are in, in that stage, and then also where, where do you see the, the future of this scenario-based planning interacting with those, uh, those other ancillary technologies?
Yeah. So currently the, the transition point is at the GTFS level of basically you do the scenario planning, you create, uh, a new network and the, and the schedule, and then you, you have, a GTFS that’ you can then go hand off to operations, uh, or, you know, feed, uh, use to elicit feedback from operations. Yeah, and I mean, that’s the beauty of the GTFS standard, GTFS static and real-time, is that, you know, tho- those are, uh, they’re, they’re standards. I mean, there’ are obviously some little nuances, and some agencies have different fields that’ aren’t a part of the, uh, the GTFS documentation, but, um, there, there are some more or less, uh, set fields, right? There are required fields that you have to be able to, uh, distribute to have a, a good working GTFS feed. One thing I would like to just add there is that I think one of the other great things is that once you have this phased implementation plan, uh, you can say, “Hey, these are the three routes that the model recommends and, and why,” and then y- or maybe it recommends Maybe you look at the top 10 and you say, “Hey, these are the three that make the most sense to us,” and then you can test them and then see, you know, where, where the predictions were right, where, where there’s maybe some tweaking in terms of, hey, there was a little bit more traffic than we expected, and, and, and being able to calibrate these models over time, uh, as, as well as calibrating some of the assumptions, then allows you to– a- a- but, but testing them with pilots allows you to then, you know, improve over time, uh, but also build trust over time.
I see. So just digging a little bit deeper there, i- is, is the, i- is, is the model learning from how you’re actually running the service, let’s say from service. change to service change? I’ve got one in January, and I’ve got one coming up for, for May, right? For like the summertime, for example.
Um, i- i- is– does it learn from service change to service change, and then you’re able to, you know, calibrate the, the model so, uh, you know, maybe you’re making those adjustments to run times and to layovers. Um, and then also maybe like year over year, how, how does that look?
And i- is it able to kinda suck in all that information and then be able to give you, uh, those precise, you know, recommendations for how to move forward?
Yeah. So wh- when we’re calibrating these models, we’re using the historical data in order to calibrate travel times, in order to calibrate ridership levels for different services for different time periods. And so as you test out more service designs and, and you– then you have more information to calibrate the model. I see. Yeah, that, that makes a lot of sense. And using the historical data, that’s just gonna be pretty beneficial for the agency too, because it’s, it’s not based on somebody else’s data. Yes. Yeah, exactly.
Uh, so, y- you know, one, one thing that, uh, I, I know that we’ve experienced at, at ETA, Anthony, is, uh, you know, just trying to figure out the workflows for transit agencies. Everyone does things, uh, you know, a little bit differently, uh, even though we all kind of have the same general understanding of what a planner does and what an o- uh, you know, an operator does, of course, a scheduler and maintenance technician.
Th- things are a little bit different, um, you know, across agencies. I- I’m curious how you’ve found th- those workflows to, uh, one, to be different in terms of doing this scenario-based planning, and then two, how do you, how do you introduce this new tool and say, “Hey, this is, this is something that you have to learn h- how to use”? And have you received any pushback on that, or people are pretty generally open to what you’re suggesting?
Yeah. There’s definitely the range of the spectrum in terms of people that are super interested and then people that are, are skeptical. Um, but I, I think a lot of the things, a lot of the skeptici- uh, skepticism, like once they see the tool, once they start using it, once they see the suggestions,
I, I think a lot of that then starts to fade away because as we explain, you know, that this is a, you know, not a black box model, it’s a deterministic model that you can rerun, it’s a repeatable workflow, we can explain how we’re predicting, uh, what riders are currently using as well as what they’re going to be using in the new system and really give them a detailed breakdown of how we think this affects, uh, both the agencies and, and the individual riders. I, I, I think that gives them a lot of flexibility in being able to interpret the results.
And, you know, you, uh, sometimes, you know, the combination of getting a, a particular schedule recommendation as well as just seeing what is the optimizer finding in terms of opportunities and where the biggest opportunities exist, um, th- there’s, there’s benefits of both of those pieces. Uh, so in our case, we’re calibrating the tr- the man model, and, and we’re predicting when people are going to be using transit or not. We’re predicting what transit they’re gonna be using, and it’s all calibrated against the transit’s historical ridership of how much ridership they’re seeing for a given route for a given hour.
And, and we’re making all our predictions and recommendations based on this, this travel demand model that’s deterministic in terms of, you know, how people are going to react to both the current network and the new network. Um, and, and so what this makes– it, it makes it very different because it’s much more like a s- a, a situation that, you know, if you were gonna use Claude to run a tool call and, and it executes code and, and then gives you– if it, if it runs the code to say, “Hey, what’s one plus two?” It’s gonna get three every time as opposed to if it says,
“Hey, what’s one plus two?” And then it looks at its probabilistic token list and then picks one of those, then it might say three, it might not say three. Wow. Yeah. That, that, that’s a big difference right there. And, you know, a, particularly for planners, we’d rather to have more consistency and, and have better understanding of the parameters. And if all parameters are the same, you know, kinda like gravitate towards the same solution, so it’s more like scientific, and it’s easier to also communicate to stakeholders, right? Because as you’re running pilots or scenarios, part of the work of the planner is not just so come up with a plan and a strategy that meets goals within budget, like, you know, a- and, and addresses all the different parameters and constraints, but also that it has the reasoning behind it, like, you know, that, that is data-driven, uh, again, to be able to present to stakeholders what should they get behind or support a given pilot and, and why are you making that effort, right? Like you’re gonna g- go with, uh, the scenarios or, or the initiatives that would yield the best results for the agency, right? That you, you know, sometimes given the circumstances, there’s, you know, s- small room for making changes that have that big impact, uh, but y- you’re gonna prioritize them and, you know, particularly, uh, go for those that have the bigger potential for improving the agency and, and yielding good results, right? Everybody wants, uh, pilots that are successful. Nobody’s looking for a pilot that is not gonna work.
Um- A- Anthony, I, I, I wanna thank you for bringing all this information a- and tell us about the future of transit planning. I, I really think it’s the future, and this technology’s gonna start transforming the way we, um, you know, plan for service in different, uh, areas around the country.
Uh, so the work that you guys do- are doing with Ontra, uh, Mobility is just remarkable, and I particularly wanna see more of how it works and what the planners see. Like, what’s that interface where they’re, uh, seeing all these different solutions? Because I think it’s, it’s, is very interesting. As we are closing our episode, we go into this, uh, segment. It’s a common segment, a recurring segment in our show that we call Rapid Fire, and we’re gonna ask you short questions, quick questions, short answers, quick answers. Are you ready for it?
I’m ready. All right. Bus or rail? I, I like both. Uh, you know, u- use both on, on a regular basis. And your model doesn’t discriminate, right? Like, whatever is the best, it’s gonna, it’s gonna recommend it. Yeah, exactly. Yeah. I, I would also add on-demand shuttles as well for, for… All three are a really important combination.
Of course. Uh, favorite transit system in the world? Um, I’m from Boston, so I gotta say the MBTA. There you go. Shout out to them.
One transit planning problem AI is well-suited to help with. Well, you know, I think service design is, is, uh, obviously the area that we’re focused on, and so I think it’s very well-suited.
But that being said, you know, not just LLMs, but using, uh, optimization, demand modeling, simulation to, to really leverage AI to solve those problems.
Okay. Best metric for ju- judging an on-demand transit pilot. I, I think it has to start with cost per ride, because I think, you know, it has to be within a certain range for it to make sense. If it’s $100 per ride for, you know, general public on-demand transit pilot, it’s not really gonna be in the questions. You gotta have it be much closer to what the buses are, are, are costing per ride, and then you can start to make it part of the conversation. Hmm.
A transit system or a agency you think is experimenting well. Yeah. I, I, I’ll give a shout-out to MARTA on this one. I feel like they’ve been really exploring and, and trying to figure out how to make things happen. I also, uh, uh, Dallas DART is one of the agencies that I would love to work with in the future. Excellent. Yeah. We see a lot of those, uh, big transit agencies.
A lot of the times they have their own innovation department and, and they’re the trailblazers and, you know, it, it… They make it a reality for all the other agencies that might not be as, um, innovative.
So yeah, shout out to those that are experimenting well out there. And the last question, rapid fire. Advice for someone who wants to work at the intersection of operations research, software, and public transit. I think with these days, with all the, all the AI tools that you have available, I think just trying out some projects and, and leveraging Codex and Claude to, to its max potential to try and, you know, experiment, uh, is, is the best way to get into it. All right. And actually, I’m just gonna throw one additional question in here. Who’s gonna win the World Cup? What does the model say, Anthony?
Uh, I, I, I’m gonna go with Spain. Okay. Okay. All right. Christian is very happy to hear that. Uh, Christian’s a, a Spain supporter. Uh, his family’s from Spain. Yes.
Anyway, all right. So, a- another recurring segment that we have here on the podcast, Anthony, is the key takeaways. I wrote down arguably too many. I’m gonna try to keep it to three, though.
Um, yeah, and one is that transit options need to compete with automobile travel times, and I, I think that y- you set that out from the very beginning as, uh, part of the problem, and I, I think that, that, that’s well-stated because, um, when you’re going to Google Maps and you’re seeing that 2X, 3X in terms of your overall travel time if you’re taking transit, that’s not gonna get anybody to, to switch to, um, a, a public transit mode like rail or bus or, or on demand. So, um, two, I wrote down that scenario-based planning is the future.
Uh, we, we need to leverage the technology to be able to do more than what just the human brain can do. Uh, and three, AI can f- help facilitate the work. It’s not something to be scared of, and it doesn’t have to be a black box.
Is there anything that you would add or change to those? Like, um, I know that they were pretty succinct, but I’m curious what your thoughts are. No, I,
I, I, I appreciate that. I, I think those three really, really landed in terms of what, what the argument was.
Excellent. Well, Anthony, this has been a blast. I’ve really enjoyed this conversation, um, and I, I think Christian too. Yes, sir. Uh, if people…
Yeah, if people want to follow along with what you’re doing at Ontra, uh, or even connect with you, uh, you know, individually, how can they do so? Yeah, they can reach me at anthony@ontramobility.com.
And I really appreciate the time, Christian and Levi. It w- it was great chatting with you. I, I loved all the questions. Excellent. Well, again, Anthony, thank you so much. Really appreciate it. This has been a, a wonderful hour we get to spend with you today. And to our listeners, thank you again for tuning in. We’ll be back next Monday with another episode of Stop
Requested.