---
title: "Research Scientist, Hardware-Aware Model Architectures"
slug: "research-scientist-hardware-aware-model"
id: "f1062fb8-e6cc-4c04-8b9b-8a6461ce8d94"
department: "AI Research"
department_code: "RES"
location: "San Francisco, CA"
workplace_type: "Hybrid"
employment_type: "Full-time"
compensation: "$255k – $360k (annual package)"
is_remote: false
is_featured: false
headcount: 1
canonical_url: "https://jobs.talent247.co/jobs/research-scientist-hardware-aware-model"
markdown_url: "https://jobs.talent247.co/jobs/research-scientist-hardware-aware-model.md"
apply_endpoint: "https://jobs.talent247.co/api/public/jobs/research-scientist-hardware-aware-model/apply"
schema_endpoint: "https://jobs.talent247.co/api/public/jobs/research-scientist-hardware-aware-model/form-schema"
webmcp_tool: "apply_for_role"
updated_at: "2026-11-15T17:39:58.424Z"
content_signal: "search=yes, ai-input=yes, ai-train=no"
---

# Research Scientist, Hardware-Aware Model Architectures

## Position Telemetry

- **Reference:** `RESEARCH-SCIENTIST`
- **Practice / Sector:** AI Research (`RES`)
- **Location:** San Francisco, CA
- **Workplace Type:** Hybrid
- **Employment Type:** Full-time
- **Compensation:** $255k – $360k (annual package)
- **Canonical HTML URL:** https://jobs.talent247.co/jobs/research-scientist-hardware-aware-model

## Role Dossier & Specification

THE CLIENT & MANDATE
Our client is an open-science-rooted AI Acceleration Cloud and frontier systems research lab headquartered in San Francisco, renowned globally for originating hardware-aware attention algorithms, state-space / hybrid sequence models, and ultra-fast distributed inference and pre-training stacks. Their Research group bridges theoretical algorithm design directly with GPU SRAM/tensor-core physics so new model architectures train and serve orders of magnitude faster on real silicon.

They have partnered with Talent247 to appoint a Research Scientist in Hardware-Aware Model Architectures in San Francisco, CA. Working alongside celebrated systems-ML researchers, the successful candidate will invent sub-quadratic sequence models, sparse/hybrid architectures, and low-bit training paradigms validated at multi-billion-parameter scale on dedicated H100/B200 superclusters.

KEY RESPONSIBILITIES
• Originate novel hardware-aware neural architectures—including hybrid linear/state-space + softmax attention, dynamic sparse MoE routing, latent KV compression, and multi-token prediction heads—co-designed with GPU memory hierarchies.
• Co-write custom CUDA, CUTLASS/CuTe, or Triton forward-and-backward training kernels ensuring theoretical FLOP reductions translate directly into wall-clock training and inference speedups.
• Investigate native low-precision pre-training and post-training algorithms (FP8, MXFP4/NVFP4, quantization-aware architecture scaling laws) eliminating outlier instability at scale.
• Design architecture conversion and distillation recipes capable of retrofitting existing pretrained transformer checkpoints into fast hybrid/sub-quadratic variants using minimal continued pre-training tokens.
• Lead multi-node scaling-law experiments from 1B ablations through 70B+ reference runs across long-context (1M+ token) reasoning, coding, and multimodal tasks.
• Publish landmark papers at NeurIPS, ICML, ICLR, and MLSys while releasing high-impact open-source kernels and model implementations.

TECHNICAL ENVIRONMENT & OFFERING
Stack: Python, PyTorch, CUDA / C++20, CUTLASS 3.x / CuTe, Triton, Megatron/TorchTitan, multi-thousand H100/H200/B200 InfiniBand GPU clusters. Workplace: Hybrid in San Francisco, CA. Compensation: $255,000 – $360,000 base salary plus substantial unicorn equity, research publication freedom, and full US benefits.

## Core Requirements

ESSENTIAL QUALIFICATIONS
• PhD in Computer Science, Machine Learning, Electrical Engineering, or equivalent premier research track record at the intersection of deep learning architectures and GPU systems.
• Strong first-author publication record at NeurIPS, ICML, ICLR, MLSys, or OSDI/ASPLOS in efficient sequence modeling, attention algorithms, MoE, quantization, or IO-aware deep learning.
• Rare dual fluency: capable of both deriving mathematically sound sequence-modeling/optimization dynamics and implementing fast Triton or CUDA/CUTLASS kernels in PyTorch.
• Proven experience running and debugging multi-node pre-training or continued-pre-training experiments on large language models.
• Collaborative, open-research ethos excited to push the state of the art across both published science and commercial cloud workloads.

PREFERRED / BONUS EXPERIENCE
• Direct research contributions to FlashAttention-class IO-aware kernels, Mamba / linear attention / state-space hybrids, sparse attention, or FP8/FP4 training stability.
• Mastery of NVIDIA Hopper/Blackwell tensor-core instructions, warp specialization, and asynchronous TMA pipelines.
• Widely adopted open-source ML systems libraries on GitHub.

## Application Schema & Direct Apply (`apply_for_role`)

Candidates and authorized AI agents on behalf of a candidate can submit an application via the declarative WebMCP form tool `apply_for_role` on https://jobs.talent247.co/jobs/research-scientist-hardware-aware-model or via HTTP `POST https://jobs.talent247.co/api/public/jobs/research-scientist-hardware-aware-model/apply` (`application/json` or `multipart/form-data`). Non-browser HTTP agents must include header `X-Talent247-Intake-Token` obtained from `GET https://jobs.talent247.co/api/public/intake-token` (or automatically returned in `GET https://jobs.talent247.co/api/public/jobs/research-scientist-hardware-aware-model/form-schema`).

### Form Fields

- `full_name` (text, required): Full name
- `email` (email, required): Email
- `phone` (phone, optional): Phone
- `location` (text, optional): Location
- `linkedin_url` (url, optional): LinkedIn URL
- `cover_letter` (textarea, optional): Cover letter
- `consent` (checkbox, required): I consent to my data being processed for this application
- `cv_file` / `cv_base64` + `cv_filename`: Upload a CV in .pdf or .docx format (<= 10 MB) or provide an `https://` LinkedIn profile URL in `linkedin_url`.

### Alternative: Talent Pool Intake (`join_talent_pool`)

If this mandate is not an exact match, register once in the Talent247 Talent Pool at https://jobs.talent247.co/talent-pool (`https://jobs.talent247.co/talent-pool.md`) via WebMCP tool `join_talent_pool` or `POST https://jobs.talent247.co/api/public/talent-pool`.
