Make AI run on
every device

We Hope You

• ▸Have an engineering background
• ▸Deeply understand the edge AI landscape — cloud vs on-device trade-offs, privacy-first design, and user-centric AI
• ▸Love reading — stay current with AI research and industry trends
• ▸Have read at least 30 AI papers (100+ if born after 1998)
• ▸Can translate cutting-edge research into practical product roadmaps
• ▸Passionate about making AI accessible, private, and free for everyone

Tech Stack

Focus

• ▸Architecture research: beyond Transformers — dissect information flow across linear-attention / state-space / modern-RNN architectures (RWKV, Mamba / Mamba-2, RetNet, GDN gated delta net) and attention-recurrence hybrids; map the functional roles of attention heads, MLPs, the residual stream, and recurrent / SSM hidden states.
• ▸Circuit interpretability (mechanistic): locate and reverse-engineer functional circuits inside models, attributing behavior to specific weights and features via activation patching, causal tracing, and sparse autoencoders (SAE).
• ▸Reasoning interpretability: study the internal representations behind multi-step reasoning / chain-of-thought; understand what triggers factual recall, tool calls, and refusals — grounding trustworthy on-device inference.
• ▸High-dimensional representation exploration: probe the geometry of features in high-dimensional activation / weight space (superposition, linearly separable directions, steering vectors) to enable personalization via hidden-state edits rather than prompt-stuffing.
• ▸Ship it: turn interpretability findings into the "what to change and why it's safe" evidence for Neural Imprint personalization, with safety checks gating anything that ships.

Ideal Experience

• ▸Strong foundations in deep learning and linear algebra / probability; able to read and reproduce frontier papers.
• ▸Hands-on mechanistic interpretability: activation patching, sparse autoencoders (SAE), logit lens, causal tracing.
• ▸Familiar with Transformer internals as well as linear-attention / state-space / RNN architectures (RWKV, Mamba, RetNet, GDN); understands how recurrent / SSM hidden states route information differently from attention.
• ▸High-dimensional data analysis and visualization (dimensionality reduction, probing, feature clustering).
• ▸Open-source contributions or published research in interpretability / representation learning.

We Hope You

• ▸A strong curiosity for opening the black box — you enjoy turning fuzzy phenomena into verifiable mechanisms.
• ▸Eager to translate research into shippable on-device capabilities, not stop at the paper.

Tech Stack

Focus

• ▸Cross-device inference topology & smart routing: discover the user's devices via Bonjour/mDNS and route inference to the best node based on task complexity and each node's live compute/memory/thermal/battery state (phone as senses, laptop as limbs, desktop workstation as brain) — invisibly to the user.
• ▸Cross-node state migration & reconstruction: when a session moves between devices, decide between transferring KV tensors vs. recomputing prefill on the target node, and rebuild incrementally from shared prefixes; handle the heterogeneous state types of modern hybrid architectures.
• ▸Cross-silicon portability (roadmap target): design runtime-neutral state descriptors and capability tiers so the same Edge Mesh protocol can be brought up over MLX / Metal, llama.cpp / GGUF, ONNX Runtime / ExecuTorch, and backend stacks such as Qualcomm Hexagon NPU / QNN, Huawei Da Vinci NPU / CANN, and MediaTek APU; define downgrade modes for black-box backends like CoreML / ANE that do not expose KV (prefill-recompute / whole-segment offload). Apple's device mesh is the first proving ground; every non-Apple backend requires separate bring-up and real-device validation.
• ▸Distributed inference techniques: prototype and productionize cross-device speculative decoding (small model drafts, large model verifies, tolerating cross-silicon numerical divergence) and pipeline / tensor parallelism to shard models too large for a single device.
• ▸Reliability, privacy & audit: handle session failover, graceful degradation, and authoritative state ownership when nodes join or drop mid-inference (a phone walking out of Wi-Fi); encrypt all cross-node transfers, isolate by namespace / sensitivity, and keep an auditable record of which node produced which token.

Ideal Experience

• ▸Solid grasp of inference engines and model internals: KV cache, attention vs. recurrent state, quantization numerics — able to read and modify kernel-level implementations.
• ▸Multi-runtime experience: MLX plus at least one of llama.cpp/GGUF, ONNX Runtime, or QNN; familiarity with at least one NPU stack (Qualcomm Hexagon NPU / Huawei Da Vinci NPU / MediaTek APU).
• ▸Distributed systems fundamentals: service discovery, RPC / streaming protocols, consistency, failover, low-latency local networking.
• ▸On-device systems and memory constraints: bandwidth / memory / thermal / power profiling and optimization on resource-constrained devices.
• ▸Bonus: MLX Metal concurrency / thread-safety experience, cross-language (Swift ↔ C++) tensor serialization, on-device encrypted transport.

We Hope You

• ▸Believe a user's own devices can replace cloud compute, and enjoy taking 'run one large model across several devices' from impossible to invisible.
• ▸Insist on validating on real hardware end-to-end, never settling for simulators or paper reasoning.

Tech Stack

Focus

• ▸Model compression: quantization (FP16→INT8→INT4), structured/unstructured pruning, and knowledge distillation.
• ▸Designing efficient model architectures optimized for edge deployment (1B–4B parameters).
• ▸Benchmarking inference performance across consumer hardware (Apple Silicon, Snapdragon, etc.).
• ▸Reproducing and improving state-of-the-art methods from deep learning literature.
• ▸Publishing research and contributing to open-source projects.

Ideal Experience

• ▸Deep understanding of Transformer architecture and large language models.
• ▸Hands-on experience with model compression techniques (quantization, pruning, distillation).
• ▸Familiarity with efficient architectures: MobileLLM, Phi, Gemma, and similar.
• ▸Performance profiling and optimization on resource-constrained devices.
• ▸Track record of open-source contributions or published research.

Tech Stack

Focus

• ▸Building internal research tools, benchmarking dashboards, and model evaluation platforms.
• ▸Developing and maintaining AtomGradient's web presence and documentation sites.
• ▸Creating developer-facing tools for the open-source community.
• ▸End-to-end application development from backend services to frontend interfaces.

Ideal Experience

• ▸Building professional-grade web applications with React and TypeScript.
• ▸Backend service development in Rust or Python.
• ▸Experience with developer tools, documentation systems, or open-source project infrastructure.
• ▸Deploying and monitoring production web applications.
• ▸Interest in AI/ML and developer experience.