You’ll own the AI models inside the filter, from training through to the build that runs on customers’ machines. You’ll decide whether a model is good enough to ship and why.
What you’ll do
- Train and improve our text classifier, which sorts pages into categories like adult, gambling and dating.
- Train and improve our image and video model, which finds nudity, skin and people so the filter can cover them.
- Cut wrong blocks on clean pages without missing real adult content.
- Fix hard cases, like modest or religious clothing marked as nudity, people seen from behind, and text written in symbols or “leetspeak”.
- Export models to ONNX and quantize them (int8/uint8, fp16). Make sure they run fast and correctly on ordinary laptop CPUs, not only new ones.
- Test each new model against the current one on real pages before handoff, and report the numbers.
- Work with the C++ team that runs the models inside the filter, so a new model drops in without breaking anything.
What we need
- 3+ years of hands-on ML work in computer vision, NLP, or both.
- Strong Python and PyTorch.
- Real experience with ONNX / ONNX Runtime and model quantization.
- Experience with object detection or segmentation (YOLO or similar).
- Experience with text classification using transformers.
- You test carefully. You check that what you ship is what you trained, and you measure before you claim an improvement.
- Good written English.
Big plus
- C++ (C++17), ideally running ONNX Runtime models from C++.
- Knowledge of CPU instruction sets such as AVX2, AVX-512 and VNNI, and how they affect quantized models.
- On-device or edge ML, where models must be small, fast and memory-safe.
- Work on content moderation or safety filtering.
- Multilingual text, especially Hebrew.