Member of ML Technical Staff

DeepGrove: Member of ML Technical Staff, Full Time

At DeepGrove (YC S25), we’re working on the next generation of efficient language models. We just raised $XXm from investors like Initialized Capital, SVA, Paul Graham, and researchers from OpenAI. We’re scaling 1 bit language models which are 10x smaller and 5-16x faster than full precision models. This enables powerful AI to be run faster, on less hardware, and entirely locally on devices on the edge.

If you’re looking for the highest ownership to model performance ratio, we are the best team to join. We’re looking to scale up our models and reach among the top US open models in the coming future.

We work across and rethink the entire modeling stack from data, optimization, infra, pretraining, post training and infra. If this sounds interesting please reach out to [email protected].

MUST HAVES:

Up-to-date with recent literature (<=1 month), papers or technical reports on the most up to date methods in area(s) of focus: frontier model landscape, infrastructure, architecture.

Strong research background or mindset (think about problems and put work in writing), e.g. technical blogs, writings, research publications.

PREFERRED:

Work on large language models at a lab (e.g. OpenAI, Google, Mistral, Z.ai, Qwen, Deepseek, Ai2, or academic).

Roles:

Pre-training:

Worked on training large scale models before or extreme obsession training of smaller models in a more academic setting.

Contributions to open community efforts in efficient pretraining: NanoGPT Speedrun, Marin, OLMo, TorchTitan, or similar.

Working knowledge of model parallelism strategies and how to combine them for a given model shape and cluster topology.

Software/hardware co-design to maximize training throughput: kernel optimization, memory and communication/computation overlap, precision choices, and interconnect-aware sharding.

Experience monitoring and maintaining long-running training jobs at scale.

Academic or open research contributions in adjacent areas: optimization and learning-rate/scaling behavior, data curation and mixing, architecture, or tokenization.

Post-training

Hands-on experience with reinforcement learning for language models — designing environments, building training infrastructure, and running large-scale training jobs.

Strong grasp of RL algorithm design and the tradeoffs between methods (e.g. PPO, GRPO, and related policy-gradient variants), including reward design, credit assignment, and stability at scale.

Experience with open-source RL training frameworks such as slime, Miles, NeMo-RL, veRL, or comparable internal systems.

Experience designing infrastructure for both synchronous and asynchronous RL, including multi-node distributed training, rollout/inference-serving integration, and weight-sync between trainers and samplers.

Academic, professional, or personal work on adjacent post-training problems: instruction and preference data curation, tool use and agentic training, evaluation design, or RLHF/RLAIF pipelines.

Infrastructure

Contributions to open-source inference engines: vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo, or comparable internal serving stacks.

Contributions to open source training repositories like Megatron, Veomni, Torchtitan, etc

Systems-level understanding of how these engines actually behave under load, and how to tune them: continuous batching and scheduling policy, KV-cache pressure and paged attention, prefill/decode disaggregation, speculative decoding, and long-context serving.

Experience with low level kernel design + DSLs

CUDA, C++, CuTE, Triton, PTX, TileLang, etc