Foundational Model Engineer — Multimodal & Agentic Medical AI Systems at SAIGroup

<p><strong>About the Role</strong></p> <p>You will be one of the earliest engineering hires responsible for&nbsp;<strong>building the technical backbone</strong>&nbsp;that powers our 3D-volume foundation model and the agentic medical AI systems built on top of it.</p> <p>This role blends&nbsp;<strong>ML systems engineering</strong>,&nbsp;<strong>high-performance computing</strong>, and&nbsp;<strong>foundation-model infrastructure</strong>, enabling our research scientists to train and deploy cutting-edge multimodal models at scale.</p> <p>You will design the pipelines, tooling, distributed systems, and evaluation frameworks that make world-class research possible—and usable in clinical settings.</p> <p>If you’re the kind of engineer who loves&nbsp;<strong>training clusters, PyTorch internals, scalable data loaders, CUDA kernels, model parallelism, and agentic inference systems</strong>, this is your role.</p> <p>&nbsp;</p> <p><strong>What You Will Work On</strong></p> <p><strong>Model Training Infrastructure &amp; Systems</strong></p> <ul> <li>Architect and maintain&nbsp;<strong>large-scale training pipelines</strong>&nbsp;for multimodal foundation models (3D volumes + text).</li> <li>Implement&nbsp;<strong>distributed training</strong>&nbsp;using data parallelism, tensor parallelism, pipeline parallelism, and FSDP/ZeRO strategies.</li> <li>Optimize training performance across&nbsp;<strong>A100/H100 clusters</strong>, including kernel-level optimizations and memory efficiency tuning.</li> </ul> <p><strong>Data &amp; Multimodal Engineering</strong></p> <ul> <li>Build scalable ingestion, preprocessing, and storage systems for&nbsp;<strong>3D medical volumes</strong>, DICOM series, voxel grids, and text datasets.</li> <li>Create multimodal data loaders and augmentation pipelines for high-throughput training.</li> <li>Work on dataset versioning, weak-label pipelines, and automatic metadata extraction.</li> </ul> <p><strong>Model Serving &amp; Agent Runtime</strong></p> <ul> <li>Build and optimize inference runtimes for&nbsp;<strong>3D-aware models</strong>&nbsp;and&nbsp;<strong>LLM-based medical agents</strong>.</li> <li>Develop robust APIs and service layers for clinical workflows (retrieval, reporting, case summarization, multi-step agent chains).</li> <li>Implement&nbsp;<strong>caching, quantization, batching, vector search</strong>, and agent orchestration.</li> </ul> <p><strong>Tooling &amp; Collaboration</strong></p> <ul> <li>Develop tools for researchers: experiment launchers, logging/visualization dashboards, model evaluation notebooks, and reproducibility tooling.</li> <li>Partner closely with scientists on&nbsp;<strong>rapid model iteration</strong>, ablations, and experimental design.</li> <li>Participate in internal “ML performance tiger teams” to squeeze maximum throughput from models and data pipelines.</li> </ul> <p>&nbsp;</p> <p><strong>Why This Role Appeals to Top-Tier ML Systems Engineers</strong></p> <ul> <li>You get to build&nbsp;<strong>the entire foundational stack</strong>&nbsp;behind frontier multimodal models.</li> <li>Rare opportunity to combine&nbsp;<strong>3D infrastructure</strong>,&nbsp;<strong>LLM agents</strong>,&nbsp;<strong>medical workflows</strong>, and&nbsp;<strong>distributed systems</strong>.</li> <li>Direct collaboration with researchers working on CLIP-style models, Chitrarth-type VLMs, document foundation models, and 3D multimodal architectures.</li> <li>Massive technical scope with freedom to propose new tools, new pipelines, new optimization strategies.</li> <li>Direct impact: your work will enable&nbsp;<strong>clinical-grade AI systems</strong>&nbsp;used in radiology and beyond.</li> </ul> <p>&nbsp;</p> <p><strong>What We’re Looking For</strong></p> <ul> <li>Strong engineering experience with&nbsp;<strong>PyTorch</strong>,&nbsp;<strong>JAX</strong>, or&nbsp;<strong>DeepSpeed</strong>, plus hands-on distributed training expertise.</li> <li>Deep understanding of&nbsp;<strong>GPU internals</strong>, CUDA kernels, NCCL, memory profiling, and high-performance data pipelines.</li> <li>Experience building&nbsp;<strong>large-scale ML pipelines</strong>, especially for multimodal or heavy-data workloads (video, 3D, imaging).</li> <li>Familiarity with cloud or on-prem HPC scheduling: Slurm, Kubernetes, Ray, etc.</li> <li>Proficiency in Python + C++/CUDA; strong command of Linux systems.</li> <li>Ability to collaborate deeply with researchers, contribute ideas, and own end-to-end engineering projects.</li> </ul> <p>&nbsp;</p> <p><strong>Nice to Have</strong></p> <ul> <li>Experience with 3D data (MRI/CT, LiDAR, voxels, meshes, NeRFs).</li> <li>Exposure to&nbsp;<strong>vector search</strong>&nbsp;(FAISS, Milvus, Annoy) and embedding retrieval systems.</li> <li>Experience with agent frameworks, LLM serving, or multimodal inference pipelines.</li> <li>Contributions to open-source ML systems or performance optimization libraries.</li> <li>Background in healthcare/medical imaging pipelines (DICOM, PACS, segmentation workflows).</li> </ul> <p>&nbsp;</p> <p><strong>What We Offer</strong></p> <ul> <li>Competitive compensation.</li> <li>World-class compute access.</li> <li>Opportunity to build the&nbsp;<strong>core infrastructure for India’s first 3D multimodal foundation model</strong>.</li> <li>Close collaboration with researchers, clinicians, and product teams.</li> <li>Autonomy, ownership, and the chance to shape the technical architecture from the ground up.</li> </ul> <p>&nbsp;</p>