Seekr builds trusted AI for mission-critical decisions. Our platform helps organizations build, govern, and deploy secure, explainable AI rooted in their own data across cloud, on-premises, edge, and air-gapped environments. We care deeply about transparency, auditability, and defensibility because high-stakes AI is only useful when people can understand and trust how it behaves.
SeekrGEO is Seekr’s geospatial AI product. This role contributes to the foundation model program behind it: pretraining and post-training of large multi-modal models on geospatial data, together with the distributed training systems that make that work possible at scale. The focus is training, but the role supports the full model lifecycle through deployment.
As a Research Engineer you lead the training systems that make ambitious model programs possible: large-scale distributed training, parallelism strategies, data infrastructure, and the operational rigor that multi-week runs demand. You will work alongside Research Scientists on modeling and recipe decisions, and you translate ideas from research papers into working code and decide whether they deserve a full training run.
Build and harden training infrastructure on accelerator clusters: data loaders, parallelism strategies, checkpointing, fault tolerance, and the evaluation harness that catches regressions before customers do.
Own the parallelism strategy for our training workloads: FSDP, tensor / pipeline / sequence parallelism, ZeRO variants, activation and gradient checkpointing, mixed precision, and the memory and throughput tradeoffs that come with each.
Diagnose distributed training failures and turn fixes into reusable platform improvements.
Design and operate the data pipeline for large training corpora: sharded formats, streaming loaders, deduplication, mixture tuning, and the versioning discipline that makes runs reproducible.
Keep multi-week training runs healthy through checkpoint management, fault-tolerant and elastic training, and the operational hygiene needed for long-horizon runs on shared infrastructure.
Do performance work on accelerators: kernel-level profiling, attention kernel selection and tuning, memory layout optimization, and closing the gap between theoretical and observed throughput.
Build the evaluation infrastructure that makes model comparisons trustworthy and reproducible, both during training and after deployment.
Support deployed models through their lifecycle: monitor systems behavior in production, diagnose regressions, and close the loop back into the next training cycle.
Contribute improvements back to SeekrFlow training so the platform gets stronger with every run.
Partner with Research Scientists to pressure-test ideas: reproduce a paper’s core claim, verify a proposed recipe scales, and turn research prototypes into production runs.
Partner with the SeekrGEO product team and customer-facing teams to align training infrastructure with the workflows the model needs to support.
Use AI coding assistants effectively as part of a modern engineering workflow while maintaining strong judgment over training code, systems code, and infrastructure.
Strong background in modern ML systems, with deep familiarity with transformer architectures, multi-modal models, and the practical realities of training them at scale.
Fluency with PyTorch and the distributed training ecosystem (FSDP, tensor / pipeline / sequence parallelism, ZeRO, checkpointing strategies).
Hands-on experience with at least one large-scale training framework such as Megatron-LM, torchtitan, or DeepSpeed.
Ability to move comfortably between engineering and research. You can read a paper, reproduce its core idea, and pressure-test whether it will hold up at scale.
Demonstrated experience contributing to a large model run through pretraining or continued pretraining, not just fine-tuning a frontier checkpoint.
Comfort designing experiments and evaluating ambiguous technical tradeoffs.
Strong Python and software engineering fundamentals, with comfort in testing, code review, CI/CD, debugging, and performance analysis.
Fluency with AI coding assistants and the modern developer workflows they enable.
Clear communication and strong collaboration across technical and non-technical partners.
Reside near Austin, TX or Reston, VA and able to work 3 days per week in office.
Experience operating distributed training at scale across accelerator clusters, with comfort in collective communication and the failure modes specific to large-scale runs.
Hands-on experience with Megatron-LM, torchtitan, and other distributed training frameworks.
Performance work on accelerators: kernel-level profiling, mixed precision, activation and gradient checkpointing, attention kernels, memory layout optimization.
Experience with AMD ROCm is a strong plus; CUDA / NVIDIA experience translates directly and is welcome.
Experience with data infrastructure for large training corpora: sharded formats, deduplication, streaming pipelines, mixture tuning.
Experience with checkpoint management, fault-tolerant and elastic training, and the operational hygiene needed for multi-week runs.
Experience with experiment tracking, model and data versioning, evaluation pipelines, and diagnosing production issues in trained models.
Track record of owning ambiguous, long-horizon technical problems, whether through graduate research, multi-year infrastructure builds, or sustained open-source or research programs.
Experience with remote sensing data pipelines and the storage or access patterns each modality demands (SAR, hyperspectral, multispectral, high-cadence EO).
Experience with infrastructure for agentic systems or tool-using models: rollouts, evaluation harnesses, RL loops at scale.
Familiarity with government and defense data handling, classification regimes, or air-gapped deployment.
Experience deploying or distilling large models for inference under real latency and cost constraints.
Open-source contributions to training stacks or geospatial ML libraries.
SeekrGEO is built for customers whose decisions carry weight. The foundation models behind that product are trained in-house, on our own infrastructure, from data we curate. This role sits at the heart of that work.
About the Company:
Company Benefits: