Jun 25, 2024
Written by Dmitry Kondratiev, Staff Engineering Manager at Bolt.
From efficiently matching passengers and drivers to finding the best way to deploy a thousand e-scooters, our data scientists are integral to shaping our competitive edge and the future of cities. We invite you behind the scenes to see how our Data Science team leverages machine learning and a relentless curiosity to make a real impact at scale:
The Data Science team currently has more than 40 people across the engineering hubs in Tallinn, Tartu, Berlin, Bucharest and Warsaw. We own any processes that require smart and automated decision-making in all areas of our business, from optimising prices to predicting routes and validating whether a scooter in a photo is correctly parked on the street. We take ownership of problems end-to-end, from discussing the business requirements with operations teams on the ground to deploying a service integrated into our backend infrastructure to receive inputs and return predictions, while providing ongoing support.
When we build our roadmap, we work with our product managers and local operations teams to determine the next most impactful thing to do. It’s not unusual for a promising product idea to originate within Data Science during a brainstorming session or a joint article review and later be included in the quarterly roadmap.
I joined Bolt after a long academic career just to find that understanding the ride-hailing marketplace is no less fun or intellectually challenging than my prior work in theoretical physics. In addition, being a data scientist at Bolt entails a whole spectrum of activities, from data/product analytics to deploying and monitoring ML pipelines. All of this, combined with an excellent company culture, makes Bolt an exciting workplace to grow professionally and personally.
David, Staff Data Scientist, Incentives
We’re focused on engaging real-life problems and creative solutions. Our data scientists come from various scientific backgrounds, making up an exciting team.
Szymon, Senior Data Scientist and Team Lead, Rides Pricing
The specific scope depends on the team’s specialisation. Let’s explore the teams currently focusing on our business’s unique areas.
The team focuses on defining fair ride pricing to meet the drivers’ and passengers’ expectations and ensuring our service is utilised to its maximum efficiency. This happens at all levels, from global optimisation to the context of a specific trip. We find the optimal base rates to grow the markets and engage dynamic pricing when demand for ride-hailing spikes, such as during rush hour or after a concert. We also ensure that drivers are well-compensated for any given order, for example, if they don’t have a chance to get a new order in the destination area. This requires a wide range of tools, from the classical supervised regression approaches to constrained optimisation and causal inference.
The team is constantly seeking the best strategies to connect drivers with passengers, ensuring reliable service and maximising completion rates. One of the team’s operational goals is to minimise pickup time, as less time spent picking up passengers allows drivers to devote more time to driving them to their destinations. We optimise the search radius for cars, ensure drivers receive orders they prefer, and support larger-scale optimisation where multiple riders and drivers are matched simultaneously.
We consider all orders on our platform with monetary incentives, like discounts for riders or bonuses for drivers, as investments. That’s an approach to business growth that’s widely applied in the industry. Our data scientists work on modelling investment strategies for managing campaign budget efficiency at micro and macro levels.
Micro-level optimisation tackles LTV, retention, churn modelling and development of novel campaign targeting methods, such as multi-armed bandits. Long-term market trends and seasonality modelling determine macro-level analysis. Here, we deal extensively with multivariate time series forecasting of various marketplace and financial indicators and multiobjective optimisation.
The Geo data science team focuses on physical space-related problems and has 3 squads focusing on different use cases.
We’re optimising the accuracy of timing and route predictions powering the dispatching engine used in ride-hailing and food delivery using a combination of a traditional graph-based router and a machine-learning layer. We work on various projects, including travel time prediction, traffic modelling, inferring missing map elements, and finding route-optimal pickup and drop-off coordinate perturbations.
Delivering food is far more challenging than it looks. Finding restaurants is tricky, delivery addresses can be hard to locate, and tight deadlines add pressure. The Delivery Experience team develops models that help couriers with last-mile delivery by identifying building entrances and parking spots while estimating courier travel time.
Our main goal is to enhance the pickup experience for our ride-hailing users. While passengers and drivers prioritise different aspects, a good pickup should generally be safe, fast, and accessible. To address this, we suggest pickup spots to riders that optimise pickup efficiency or, if possible, make their ride a bit cheaper. Considering factors like driver availability and local traffic conditions, we aim to reduce friction and wasted time during pickups, making our service more efficient.
The team is responsible for all data science tasks for Bolt Food, which connects eaters and food providers like restaurants or stores. The team handles provider ranking in the Bolt Food app, textual search, user food recommendations, item image and text generation, dynamic pricing for delivery, dispatch optimisation, delivery and cooking time prediction, and geographical courier distribution. The team also supports the Bolt Market network of dark stores (local warehouses for online shopping). Here, we tackle the supply chain optimisation, user search experience and item pricing.
The team develops tooling to verify the identity of Bolt users and drivers by checking the data from the documents they provide. Afterwards, we match the face of the person to the document and make sure that both the face and the document are first-hand captures belonging to the real person. We use vendors’ and in-house text extraction and liveness detection systems to provide state-of-the-art performance and scalability.
The team helps optimise various routines within Bolt. Whether assisting our Customer Support team in handling incoming requests more efficiently or enabling our mapping specialists to improve our routes and maps, the team has the right tools for the job. We handle diverse projects, including customer ticket classification, workload forecasting, service-level optimisation, and end-to-end customer support automation. This ranges from the creation of a support ticket to offering automated solutions. We employ various tools, from traditional machine learning models to Large Language Models (LLMs), to build agent-like systems.
We’re solving the challenges of running micromobility and car-sharing services at scale. One of our core streams is safety and compliance. We leverage sensors and camera data to detect dangerous riding and lousy parking. On top of that, we’re improving operational efficiency through predictive maintenance of our scooters and using levers like pricing and deployment allocation to optimise fleet utilisation. Consequently, we’re a diverse team interested in IoT, optimisation, and deep learning.
The Fraud team leverages machine learning models to improve fraud detection effectiveness: catch fraudsters without affecting legitimate users across all Bolt business lines. Additionally, the team helps scale fraud operations by increasing the number of automated decisions, thereby enhancing efficiency and supporting sustainable expansion.
Our core fraud ML models address payment-related fraud, including chargebacks and failed payments. We face challenges like dealing with highly imbalanced fraud data and frequently changing fraud labels. To ensure our models remain effective, we prioritise continuous experimentation, measuring the impact of model updates to refine and enhance our fraud detection capabilities.
We empower product teams at Bolt to run A/B tests and make better product and business decisions by providing self-service tools and guidance. We’ve built a scalable, cost-efficient, end-to-end in-house experimentation platform with the support of advanced statistical methods, switchback tests, and custom metrics to support diverse use cases from all business verticals inside Bolt.
We work on detecting incidents and safety events involving both drivers and riders. We aim to predict and prevent any incidents or, when that’s not achievable, detect them as soon as possible so we can take proper actions to mitigate them. Most of our challenges lie in detecting rare events across heterogeneous markets and environments. We have to balance the cost of false detections, be that customer support costs or degraded user experience, against the increase in the safety of our users and partners.
Working in the Data Science team is incredibly rewarding because our models make a real difference for millions of passengers and drivers. I particularly enjoy the opportunity to amplify this impact by developing a new stateful streaming processing system that tackles the complex task of performing stateful streaming operations, all to deliver real-time information to our models. The abundance of technical challenges, combined with the trust and autonomy we’re given, allows me to innovate and grow.
Andrzej, Machine Learning Operations Engineer, Rides Pricing
Python and its machine learning toolkit are the primary instruments a data scientist uses at Bolt. There are no requirements to use a particular library, as different tools may fit different objectives. And as you saw above, we have a wide array of tasks to solve. However, we developed some internal tooling around the libraries we use most often to reduce the boilerplate effort of a new project.
We use SQL to get the data we need to train the models. We use Presto and Spark engines to access the data for daily ad-hoc queries. Still, all production loads rely on the latter due to its efficiency in processing big data and flexibility, from fine-tuning the cluster settings to making custom Python UDFs.
The end product of most data science projects is a model deployed in a container in Bolt’s internal infrastructure, ready to communicate with the backend services to receive real-time calls and return predictions. The most loaded models are serving hundreds of thousands of requests per minute. A dedicated Model Lifecycle team maintains a framework that abstracts most infrastructure issues away from the daily data scientist work to smoothen the process. You need to write only a minimal amount of code to have your model automatically retrained, tested and redeployed on a schedule. They also help us with scaling (for example, for New Year’s Night when our infrastructure experiences the most load) and controlling the cost.
At Bolt, I’m surrounded by exceptional colleagues whose kindness and intelligence create a supportive work environment. On a professional level, I enjoy the complexity of the systems and Bolt’s commitment to modern technologies. I learn new things every day.
Sophie, Data Scientist, Geo
I’m thrilled to have joined a company dedicated to solving one of the most pressing issues of our time: finding better ways to move around in our cities and reducing our reliance on stressful and polluting personal vehicles. Not only that, but I also get to work in a dynamic, creative, and enjoyable environment, collaborating with talented and supportive colleagues from around the world. I’ve participated in ping-pong matches, augmented reality puzzles with my local team, online games with remote colleagues, and even a DJ set on our terrace with a view of the iconic TV Tower in Alexanderplatz!
Pablo, Senior Data Scientist, Rides Matching
Even though different data science teams focus on different problem areas, we remain together as a professional data science community within Bolt. We host regular workshops with other teams, where we can learn about the latest advancements and use cases from our colleagues and share and receive feedback on the latest projects. Every data scientist is encouraged to contribute to common tooling that helps us be faster and more efficient.
To develop our expertise as a team, we share the latest tips and tricks in joint sessions, run brainstorming sessions to reuse each other’s latest advancements and hold joint reading sessions to review the latest relevant literature. In addition, Bolt helps us fund the books and courses so we can stay on top of the latest developments or explore the topics we’re currently working on in more depth.
Rotations between different data science teams are common after a reasonable tenure in one’s current team, and it helps us to keep the teams’ vision fresh and satisfy the natural desire to learn new things. After all, as data scientists, we’re all curious to explore, and a fresh perspective always helps.
The first significant milestone in a data scientist’s career is reaching Senior grade. That’s unless you wish to switch tracks and try different careers like analytics or software engineering. However, people usually switch to data science from these domains instead. After you become a Senior, where you have developed advanced expertise in data science and business thinking and shown that you can drive bigger projects independently, there are two paths forward: technical leadership or people management.
There are 3 main ways in which Data Science in Bolt differs from other companies, which I really appreciate:
DS has a real influence on product decisions. Thanks to a relatively flat hierarchy, Engineering, Product, and DS have a strong position in identifying the opportunities for product development. It allows us to shape the way a product would work. Instead of working with strict requirements coming from above, there is flexibility in generating ideas and solutions by yourself.
Lots of strong colleagues. There is a strong DS community inside Bolt, with many people with different backgrounds and skill sets. For Individual Contributors, it opens a possibility to quickly grow, learn best practices and discuss new models and frameworks.
Task diversity. Bolt Data Scientists work with everything, from classical ETA predictions, classifications, and fraud detections to pricing, computer vision, and generative AI topics. One can always find an interesting area and work there. Additionally, Bolt is flexible in allowing transfers between DS sub-teams, so you don’t get stuck on a single task.
Oleksandr, Senior Data Scientist, Bolt Food
At Bolt, we’re eager to welcome data science professionals looking for a vibrant community and dynamic challenges. Join us and see how your work shapes the future of cities. And check out our video to get even more inspired.
.png)
