Hybrid – 3 days onsite.
The ML Engineer is a new role , responsible for shaping how we build and scale machine learning systems at AP, helping to lay the foundation for our machine learning capabilities. The ML Engineer has hands-on experience building and optimizing ML inference systems that run in production environments. This role will develop and tune pipelines that transform millions of photos, videos, and text documents into searchable representations using a combination of deep learning models (examples: DistilBERT, SBERT, TransNetV2) and external multimodal APIs. The ideal candidate has experience optimizing inference at scale, orchestrating ML workloads, and working with both PyTorch and TensorFlow in a cloud environment, focusing on model performance, integration patterns, and inference efficiency.
What You’ll Do:
· Design, build, and scale ML-powered inference systems that process large volumes of text, image, and video data to power news-based intelligence products.
· Productionize and optimize state of the art models and inference pipelines. These models include, but are not limited to:
o DistilBERT for Named Entity Recognition (NER) over hundreds of thousands of search queries/day
o TransNetV2 for video shot boundary detection at scale for archival video as well as real-time
o SBERT for embedding generation from textual descriptions
o External multimodal APIs for image/video captioning
· Support hybrid search architectures by defining embedding/re-ranking interfaces, evaluation metrics, and inference performance requirements; partner with search/platform engineers on index configuration, sharding, and cluster tuning.
· Design and implement scalable data processing pipelines across hybrid CPU/GPU environments to handle millions of media assets.
· Partner with MLOps and platform engineering to enable the deployment and operation of ML systems reliably, contributing to:
o Distributed inference architectures
o Cloud-based execution (e.g., AWS EC2, Batch, Lambda, SageMaker)
o Efficient resource utilization across workloads
· Optimize inference latency and throughput across distributed workloads using cloud-based resources (AWS EC2, Batch, Lambda, SageMaker, etc.)
· Build resilient asynchronous processing systems for large-scale workloads, ensuring:
o Reliability (retries, fault tolerance)
o Efficiency (caching, deduplication)
o Observability (metrics, logging, traceability)
· Work closely with data scientists and product teams to iterate on models, improve performance, and deliver measurable impact in production.
Requirements:
· 8+ years of experience building production ML inference systems.
· Demonstrated ownership of deep-learning inference optimization in production (quantization, distillation, compilation, kernel/profile-level performance work) for transformer NLP and/or CV models.
· Experience with both TensorFlow (SavedModel, tf.data, XLA, TFLite) and PyTorch (TorchScript, ONNX, FastAPI/TorchServe)
· Hands-on experience optimizing inference pipelines on AWS infrastructure, ideally across different types of media assets.
· Experience with video frameworks/tools (e.g., FFmpeg), and working with large-scale frame-level inference.
· Demonstrated experience monitoring and debugging model latency, memory, and pipeline throughput.
· Experience with hybrid search architectures (BM25 + vector search + cross-encoder reranking).
· Familiarity with OpenAI APIs or other foundation model providers.
· Familiarity with open source HuggingFace LLMs.
· Experience with data pipeline and workflow orchestration tools (e.g., Airflow)
Who This Role is Not For:
Candidates whose primary background is MLOps platform work (DAG orchestration, Terraform, Kubernetes administration, generic CI/CD pipelines) will not be a fit. We need a senior level engineer who can profile a transformer, rewrite its serving path for a 2–3x latency reduction, tune an HNSW index, and tell us which SageMaker instance type will hit our p95 target at the lowest cost
Never miss an opportunity! Create an alert based on the job you applied for.
