Machine Learning Engineer

Blissway is a startup that simplifies toll collection and improves highway safety through deep technology, AI/ML, hardware, SaaS, and IoT. The Machine Learning Engineer will own the end-to-end machine learning pipeline, from raw sensor data and dataset curation through production deployment, monitoring, and iteration across cloud and edge systems. The role focuses on computer vision and shipping models for real-world roadside hardware and large-scale image processing.

Responsibilities:

• Own the Whole Pipeline: You decide which problems are worth solving, then take them from raw sensor data all the way to production: collection, dataset curation, training, deployment, monitoring, and iteration. We wire it all together and run our own servers, so the pipeline is yours end to end
• Real Hardware in the Real World: This is the part that makes us special. We own the devices in the field. This means any idea you have can actually get built and tested on real roads
• Vision at Real Scale: We process 11 million images every single day, running detection, segmentation, classification, embeddings, and re-identification across everything in the frame: vehicles, license plates, wheels, even lane markings. Beyond images, we have multiple other sensors on the road pulling in different data making the problem space wide open. At this volume, the right model can drastically improve accuracy and cut cost at the same time
• Classical CV to Custom SOTA: Our toolbox spans the full range, from traditional computer vision algorithms to custom-trained state-of-the-art models (detection, segmentation, embeddings, classifiers). You pick the right tool, and when nothing off-the-shelf is good enough, you train your own
• Build the Best Models That Exist: We read the papers, go to the conferences, and hold our work to the current frontier. ML has become essential to Blissway over the past year, and this team is where that bet gets made real
• Edge and Cloud: Most of our compute lives in the cloud where power is effectively unlimited. We're now pushing more inference onto the roadside hardware itself, a completely different problem: the models have to be fast, small, and power-efficient without giving up accuracy. You'll work both sides of that constraint