3 months ago
San Cugat del Vallés, SpainSenior
Responsibilities
- Own the architecture, representation, training, and improvement of models that read handwritten pen strokes.
- Design and implement approaches for multiple-instance detection over variable, non-rasterized geometric primitives.
- Run controlled comparisons with predefined success criteria and analyze detailed regressions and improvements.
- Replicate intended results, document unsuccessful experiments, and critically evaluate research literature.
- Deploy models to ordinary school hardware with fast inference, matching evaluation behavior, determinism, and rollback capability.
- Monitor live performance on student work and use observed failures to guide subsequent experiments.
- Collaborate with dataset and evaluation owners to identify whether limitations arise from the model, data, or task ambiguity.
Requirements
- Deep experience with multiple-instance detection or instance segmentation, including matching predictions to ground truth and interpreting relevant metrics.
- Experience with non-rasterized inputs such as strokes, polylines, trajectories, point clouds, graphs, vector graphics, CAD geometry, or sensor traces, or a clear understanding of the required modeling changes.
- Experience with attention-based architectures over sets or graphs and encoding spatial and relational structure.
- Serious hands-on experience training deep models in PyTorch, including training schedules, numerical stability, distributed runs, and debugging.
- Evidence of taking models into production, including export, latency, determinism, and rollback.
- Strong experimental discipline and the ability to distinguish meaningful improvements from noise or benchmark artifacts.
- Ability to read technical literature critically and reproduce useful methods.
- Experience with online handwriting, sketches, ink, diagram recognition, geometric deep learning, graph neural networks, vector graphics, CAD, GIS, trajectory modeling, inference optimization, or quantization is preferred.
- Published or open-source work and Spanish or Catalan are nice-to-have qualifications.
Benefits
- Access to the GPUs needed for experimentation.
- Small technical team with direct access to product decision-makers and substantial ownership of model direction.
Tech Stack
Categories
