Gravel is an imprecise term and ideal tire-size debates on the internet for a given route can go deep in the comments section. Slipstream AI is a new route-planning tool that hopes to bring a little more objectivity to how we rate gravel technicality for drop-bar riding. Read on to learn more about how this application is leveraging artificial intelligence to improve gravel cyclists’ experiences.

Raise your hand if you’ve ever gotten caught unintentionally underbiking on a new-to-you route. Raise your hand if you’ve ever been late to dinner, missed your resupply’s operating hours, or have generally been sandbagged on an unexpected stretch of hike-a-bike while riding unfamiliar terrain. If you ride drop-bar bikes on the all-encompassing and poorly defined surface of gravel, you’re probably holding your hand in the air right now (or, at least nodding along, knowingly).

Most mapping platforms used by cyclists—like Ride with GPS, Strava, or Komoot—use fairly limited coding keys to visually communicate a route’s surface, making it difficult to discern the level of technicality beyond a binary dirt-or-pavement distinction. In an app’s route-planning mode, pavement might be delineated as a solid line, while anything that’s not pavement gets a dashed-line treatment. Mountain bike trails may be shown as thinner, darker-colored lines in the original map layer, but for a novice route-planner, there’s nothing that keeps this more advanced terrain from being added to a planned drop-bar route in a few clicks and then taking on the appearance of just more dashed-line “gravel” territory.

Gravel (left); also gravel (right)

Gravel is the umbrella term for all drop-bar friendly off-pavement terrain, but anyone who has sampled the discipline extensively knows that not all gravel is created equal, despite what the lines on the map may show. So how do you know if those dashed lines are going to translate to champagne dirt, or bumbley babyheads? Slipstream AI is a new platform that’s trying to make it easier for gravel cyclists to read between those mapped lines.

Slipstream AI: Backstory

Slipstream AI is a mapping platform that uses rider-submitted images to generate technical ratings for mixed-surface roads and trails. It was founded in April 2024 by Boulder, Colorado, locals Alex Haeger and Jenna Blumenfeld. In a recent conversation, Alex told me that, early in his and Jenna’s relationship, he’d gotten sandbagged while out riding before a planned dinner date. The experience may have been early inspiration for the solution that Slipstream now offers.

“You’re up in the mountains, way out, and the section of road you thought would take you an hour is now taking you three hours, because you’re hiking, and you can’t send your date a text because there’s no service,” described Alex of the all-too-relatable experience. “We all have stuff to do and it’s quite important how we budget our time.”

Fortunately, Alex being late for the date didn’t prove to be a dealbreaker for the pair, but later, while bikepacking together in Spain, they finally decided that they were done getting skunked by slower-than-expected terrain.

“I had spent a lot of time building a route from Barcelona to where we wanted to ride to visit friends,” recalled Alex. “We’re on this perfect gravel road when the headunit tells us to go right. We look right and there’s no trail—classic example of ‘Oh the existing mapping platforms think this is a trail but it’s really not.’”

The challenges of that trip were also still fresh for Jenna; “We ran out of food multiple times in Spain due to a lack of road surface information. In our third hour of hike-a-biking through Catalonia, eating only sugar packets, the idea for Slipstream AI started to materialize.”

Slipstream AI: Building a Model That Rates Gravel

At that point, frustrated by a lack of definitive available information, or an accessible outlet to share information on, the couple decided to try building their own tool for gravel cyclists and bikepackers. Their goal was to create a more objective categorization system for road surfaces—known as the Slipstream “Road Surface Rating” (RSR, pronounced “razor”)—that would better inform riders’ expectations.

“The descriptors we have around gravel riding are imprecise,” said Jenna. “What’s ‘gnarly’ to one person may be a walk in the park to another. With Slipstream AI’s standardized RSR rating, we have a common language to describe what we’re riding.”

With a background in environmental engineering, Alex had picked up enough coding knowledge from his work to build the AI model himself. It took six months to develop the first iteration but, as Alex said, “I was between jobs on unemployment. I had some time and I really wanted to make this thing happen.”

He said that amassing enough data to train the model on Slipstream’s 1-to-6 RSR scale was the hardest part, “believe it or not, there’s not a lot of photography available on the internet of gravel roads and trails.” Training the AI was kind of like studying for a test with flashcards, a process that involved showing the model an example photo of each Road Surface Rating classification (“This is pavement, this is smooth gravel,” etc.) and repeating that process a mere 30,000 times. The training process mirrors how the AI will process user-submitted photos: photo in, RSR out.

Slipsteam AI’s Road Surface Rating Scale

• RSR 1: Paved roads and bike paths. Concrete, asphalt, pavers.
• RSR 2: Smooth gravel open roads. These are the clean gravel roads we all dream of.
• RSR 3: Uneven and potentially rutted gravel and dirt roads.
• RSR 4: Gravel/dirt singletrack that is free of rocks.
• RSR 5: Uneven, rutted, and rocky gravel roads.
• RSR 6: Rocky gravel singletrack (think: blue mountain bike trail, sections of embedded rock. About the upper limit of what one could get away with on a drop-bar bike).

Slipstream defines the RSR scale loosely around what one could ride, or feasibly get away with riding, on a drop-bar bike. In creating the scale, Alex said he tried to factor in some of the things he learned in civil engineering about how roads are maintained. “Generally, a 2 is going to be a 2 every year, almost regardless of season because the city/county are going to take care of it,” he explained. “A 3 will almost always be a 3—it could be a Minimum Maintenance Road, or a country two-track—they’re usually narrower, uneven and the edges are less defined.”

“We call it an RSR because it sounds cool,” joked Alex (see: the pre-historic flip-phone and resurgent scooters by the same name), but he hopes the rating system will have serious and positive impacts on the world-wide cycling community. “The technical terrain can change so much on a gravel route and it’s not as inconsequential as it might be on a mountain bike.”

Slipstream AI: How It Works

Today, Slipstream is available for free for cyclists around the world to use. The site is straightforward, with the RSR color-coded map being the main feature. Pavement (RSR 1) gets coded in green, and the coding intuitively moves up, through yellow, orange, and red (RSR 6) based on difficulty. The only sections of the map that are color coded are those areas for which users have submitted images through the site’s uploader tool. More images for a given stretch of road result in denser and more continuous lines of color-coding; less images for a road will show as single dots, or data points, which is essentially what they are—a snapshot of what riders will experience on the ground.

For data liability purposes, cyclists can only access the upload tool on the platform by making an account (though the ever-evolving map is visible to all). Slipstream says they worked with trusted, vetted tech advisors to establish database security and data privacy protocols to guard users images (for the images to work with the platform, they must have a GPS tag—this is built into the file structure for iPhone and GoPro photos). Images uploaded to the platform are never used publicly by Slipstream, but stored in deep server storage for the purpose of further training the model. Once an account is made, cyclists can upload their own photos to the database to continue filling in the RSR map.

That, in fact, is the primary hope of the platform: to create a wealth of user-generated data that then improves users’ experience while out riding. Alex and Jenna say that they don’t want Slipstream to “guess” about blank sections on their map and risk presenting potentially misleading information, like what they encountered on their trip in Spain. Alex said he can’t be completely sure of where other companies are sourcing their data, but he has some ideas:

“Absent creating the data like we’re trying to do, if you’re scraping the internet you’re really limited to just a few areas. One being municipal databases that might say a road is paved, but that road might go all the way up a mountain and, at some point, turn to gravel, but the database may only say that it’s paved. Or, you can scrape from places like Garmin, where people create rides that have been labeled as [all] Gravel routes, and you have thousands and thousands of these routes, so you’re kind of triangulating in a way to get towards the data. But it’s never actual data.

“We’re trying to take a different approach because we know people want better information, if anything just to plan your rides. I tried scraping in the same way that [other companies] do and there’s just no way to get reliable information. We want to provide trustworthy information without going so far as these other platforms that are just kind of guessing [on missing sections]. If there’s a mile stretch of road without data, we’re not going to guess about it. But at least people will be able to see what’s before and after it.”

Based on side-by-side comparisons of what the model has generated and verifiable road surfaces, Slipstream’s AI is quite strong, yielding results with 98% accuracy at this point. According to Alex, “You can be pretty sure that if you send your photo(s) in, it’s going to get rated accurately. We want to make sure that if people are taking the time to send us photos, then we’re doing what we can to make sure it gets coded correctly on the map right.” (The one caveat here is sand, which Alex admitted is tricky to code. Because of its lack of visible features like rocks, sand is typically giving an RSR of 3.)

Slipstream AI: Precision = Confidence

Slipstream’s recent move to offer course previews for races is one way that the company is hoping to continue expanding its “high fidelity” data, while also offering a valuable service to riders who might not have time to pre-ride an entire course. As part of a paid partnership with an event, Alex and Jenna will ride each distance for a race with a GoPro mounted on their handlebars, set to automatically click off a photo every 10, 20 or 30 seconds, depending on terrain. Once the photos are uploaded to the platform, a granular preview is created on Slipstream’s map for each course. (Slipstream recently made branded GoPro/Garmin mounts that they’re sending out for free to individuals who want to help gather data in this way to continue building out RSR coding in their area).

After the 2024 edition of Ned Gravel, one of Slipstream’s Colorado race partners, Alex and Jenna heard firsthand from participants about the benefits of the preview. “There was one stretch of road that’s a sweeping downhill into a hard left turn,” said Alex. “It goes from [RSR] 3 to 5 right before the hairpin and people were coming up to us after the race and were like ‘Thank you so much! We studied the route and we saw that RSR 5 out there and we kept our eyes open for it.’ It was cool to hear how it helped people prepare for the ride.”

For Jenna, this real-world example is proof of concept of the confidence that they’re hoping to inspire with Slipstream; “We created Slipstream AI because surface quality data is empowering! We hope to provide valuable insights to help you ride with confidence.”

Looking Ahead

Less than a year since it launched, today Slipstream has added data to the RSR map from over 100,000 user-uploaded images, and has seen users from 79 countries. It makes sense that the map’s densest data population is in the Front Range around Boulder, but Alex and Jenna are hopeful that it will continue to fill out the world over. “People are just excited to share and excited to explore,” said Alex.

Although both Alex and Jenna have careers outside of Slipstream, they’re working hard at building additional features to roll-out on the app. Most recently, they’ve launched a Strava integration that breaks down actual percentage of road versus gravel, and each subsequent RSR categories, for a given route in a rider’s Strava caption. And they’re clearly having fun with it. “Some days the AI will talk like a pirate just to keep people on their toes,” laughed Alex. “We want to do other things like that to spice things up.”

Further down the line, they’re working on an integrated route-building feature as, in addition to single images, currently Slipstream can only rate GPX routes submitted from other platforms. Aside from removing a step for users in the process of route planning, incorporating their own mapping tool would create a compelling interplay between the RSR scale and elevation profiles (e.g., for a steep, downhill RSR 5, you know you’d have to watch out).

Given that they each work full-time outside of Slipstream, Alex and Jenna view making money from the venture as a low priority. Right now, Slipstream has incredibly low operating costs and is fairly hands-off by design. In a world where both tech and gravel seem to have strayed from more grassroots origins, it’s refreshing to see the two combined again in a DIY-level platform that still has the potential for mass appeal. When asked what success would look like though, Alex isn’t afraid to dream big, “We want to be the only platform that gravel cyclists need to share information, get information and have a social outlet for gravel riding.”

See more at Slipstream AI.