The RoleSpecter is hiring an applied AI software engineer to own the path from model to product. You'll take detection, tracking, and vision-language models from research prototypes to reliable production systems running across a distributed fleet of edge devices and cloud inference services, then close the loop by measuring how those models actually perform in the field and feeding that signal back into the next iteration. This role sits between our research team and our product surfaces, owning the evaluation infrastructure, data pipelines, and deployment mechanics that determine whether a model improvement becomes a customer-visible improvement.
Responsibilities:- Productionize models from the research team, spanning containerization, inference optimization, and deployment to edge devices and cloud GPU infrastructure.
- Build and own offline and online evaluation systems that measure model quality against production traffic, not just benchmark datasets.
- Design data feedback loops that surface failure cases, route them to labeling, and turn them into training and evaluation sets.
- Define and instrument the metrics that matter for perception quality in deployment (false positives per camera per day, identity switch rate, latency budgets, drift over time).
- Ship product features that depend on new model capabilities, working end to end from inference service to API to the behavior a customer sees.
- Run experiments and staged rollouts across the fleet, including shadow deployments and A/B comparisons between model versions.
- Manage inference cost and performance tradeoffs across hardware targets, from constrained edge compute to serverless GPU providers.
- Collaborate with research and platform to make model iteration a routine deployment rather than a bespoke project.
Qualifications:- Experience taking ML models from research code to production systems that real users depend on.
- Strong Python skills, with working knowledge of PyTorch and modern inference runtimes (ONNX, TensorRT, or similar).
- Track record building evaluation pipelines and defining metrics for model performance in deployment, including handling distribution shift and unlabeled production data.
- Experience with data pipelines and annotation workflows for computer vision or other perception domains.
- Comfort with infrastructure: Docker, Kubernetes, cloud GPU services, and CI/CD for model artifacts.
- Familiarity with computer vision (object detection, multi-object tracking, re-identification) or vision-language models.
- Experience with streaming or event-driven data systems (Kafka, Redpanda, or equivalent).
- Product instinct - you can reason about which model improvements will actually change what a customer experiences.
- Interest in physical AI, IoT, or large-scale distributed sensing systems.
