ai& Company Job Analysis

Created: 2026-03-25
#jobs#ai-infrastructure#career#gpu#kernel-optimization

ai& Company Job Analysis

Company Overview

ai& is a global AI technology company focused on:

  1. AI Lab: Specializing in localization
  2. Infrastructure Provider: Global compute and data centers
  3. Full Stack Ownership: Believes in owning the stack top-to-bottom

Locations: Tokyo, SF, Austin, Toronto (global hiring) Culture: Small autonomous teams, high personal agency, mission-driven

Job Roles Analysis

1. Member of Technical Staff - Inference Optimization

Focus: Extract maximum performance from GPU hardware

Key Responsibilities:

  • Write custom CUDA and ROCm/HIP kernels
  • Profile with Nsight Compute, rocprof, Perfetto, VTune
  • Optimize attention mechanisms (Flash Attention, MLA, paged attention)
  • Work with quantization (INT8, FP8, AWQ, GPTQ)
  • Fuse operators to reduce memory round-trips
  • Optimize collective communication (AllReduce, AllGather)

Required Technical Skills:

  • Deep CUDA/ROCm kernel authorship
  • GPU architecture knowledge (Hopper/Ampere, CDNA)
  • Profiling tool fluency
  • Numerical precision expertise (FP16, BF16, FP8)
  • Parallel programming (CUDA, Triton, SYCL)

Your Learning Path:

  • Phase 1-5 of GPU learning path
  • Focus: CuTe/CUTLASS for peak performance
  • Critical: Profiling and bottleneck analysis

Difficulty for Your Background: ⭐⭐⭐⭐⭐ Very Hard Time to Competency: 12-18 months Priority: High - Core technical role

2. Member of Technical Staff - Inference Serving

Focus: Build high-performance multi-tenant serving stack

Key Responsibilities:

  • Evaluate and tune inference frameworks (vLLM, TensorRT-LLM, etc.)
  • Optimize latency (TTFT/ITL) vs throughput trade-offs
  • Implement disaggregated prefill/decode, speculative decoding
  • Manage KV-cache for multi-turn conversations
  • Support Day 0 model releases
  • Bridge small clusters to massive multi-node deployments

Required Technical Skills:

  • Deep experience with inference engines
  • Understanding of "Inference Triangle" (latency/throughput/quality)
  • Multimodal model serving (Video, Vision)
  • Distributed systems at scale

Your Learning Path:

  • Phase 1-3 of GPU learning path
  • Focus: Triton, PyTorch internals
  • Study: vLLM, continuous batching, paged attention

Difficulty for Your Background: ⭐⭐⭐⭐ Hard Time to Competency: 9-12 months Priority: High - Close to kernel optimization

3. Member of Technical Staff - Systems

Focus: Bare-metal infrastructure and GPU systems

Key Responsibilities:

  • Configure and manage bare-metal servers
  • OS tuning, driver management, firmware
  • Rack-scale GPU systems (NVL72, AMD Helios)
  • GPU topology (NVLink, NVSwitch, MI-series)
  • Performance optimization at system level

Required Technical Skills:

  • Linux internals, kernel parameters
  • NUMA topology, PCIe configurations
  • CUDA driver/firmware management
  • High-density GPU system operations

Your Learning Path:

  • Phase 1 of GPU learning path
  • Linux system administration
  • Hardware architecture study

Difficulty for Your Background: ⭐⭐⭐ Medium Time to Competency: 6-9 months Priority: Medium - Good entry point

Why This Fits You:

  • Leverages systems thinking from control engineering
  • Less deep kernel programming required
  • Strong foundation for moving to optimization later

4. Member of Technical Staff - Networking

Focus: Lossless networking fabrics for GPU clusters

Key Responsibilities:

  • Design RoCE v2 and InfiniBand fabrics
  • Tune NCCL for distributed AI workloads
  • Optimize collective communications (AllReduce, AllGather)
  • Network topology design (fat-tree, rail-optimized)
  • Congestion control and zero packet loss

Required Technical Skills:

  • AI networking expertise (InfiniBand, RoCE v2)
  • NCCL tuning
  • NIC/DPU configuration (ConnectX, Bluefield)
  • Network architecture at scale

Your Learning Path:

  • Networking fundamentals
  • RDMA and InfiniBand concepts
  • NCCL and collective communication patterns
  • GPU topology integration

Difficulty for Your Background: ⭐⭐⭐⭐ Hard Time to Competency: 9-12 months Priority: Medium - Specialized but critical

Why This Fits You:

  • Control systems background → network control theory
  • Systems thinking applies to network optimization
  • Less code-heavy than kernel optimization

5. Member of Technical Staff - Post Training

Focus: RL and post-training pipelines

Key Responsibilities:

  • Build post-training infrastructure (SFT, preference alignment)
  • Run RL training experiments
  • Design synthetic data pipelines
  • Enterprise customer post-training engagements
  • Continual learning without catastrophic forgetting

Required Technical Skills:

  • RL implementation on language models
  • PyTorch, DeepSpeed, FSDP, vLLM
  • Data generation and quality assessment
  • Evaluation pipeline design

Your Learning Path:

  • RL fundamentals (PPO, reward models)
  • PyTorch training infrastructure
  • Less focus on low-level GPU kernels

Difficulty for Your Background: ⭐⭐⭐ Medium Time to Competency: 6-9 months Priority: Medium - ML focus, less systems

6. Member of Technical Staff - Applied & Agentic AI

Focus: Build agentic layer and frameworks

Key Responsibilities:

  • Integrate agentic frameworks (LangGraph, AutoGen, MCP)
  • Build deployment tooling for agents
  • Design custom agentic workflows
  • Model evaluation and QA
  • Developer experience for agent deployment

Required Technical Skills:

  • Agentic framework experience
  • Applied ML fluency (prompting, chaining)
  • Evaluation pipeline design
  • Full-stack development
  • Python, agentic frameworks

Your Learning Path:

  • Agentic AI frameworks
  • LangGraph, AutoGen tutorials
  • Less GPU kernel programming

Difficulty for Your Background: ⭐⭐ Easy Time to Competency: 3-6 months Priority: Low - Not aligned with kernel optimization goals

Note: This is furthest from your GPU learning path but easiest transition.

7. Member of Technical Staff - Applied AI Engineer (Internal Tools)

Focus: AI-powered internal tooling

Key Responsibilities:

  • Build AI-powered internal tools
  • Deploy agentic workflows internally
  • Cross-functional automation (finance, ops, HR)
  • Tool evaluation and iteration

Required Technical Skills:

  • Applied AI engineering
  • Full-stack development
  • Agentic framework fluency
  • Cross-functional collaboration

Your Learning Path:

  • Similar to Applied & Agentic AI role
  • Focus on internal tooling patterns

Difficulty for Your Background: ⭐⭐ Easy Time to Competency: 3-6 months Priority: Low - Not aligned with kernel optimization goals

Strategic Analysis

Best Roles for Your Goals

If you want to learn GPU/kernel optimization:

Rank Role Relevance Difficulty Why
1 Inference Optimization ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐⭐ Direct kernel programming
2 Inference Serving ⭐⭐⭐⭐ ⭐⭐⭐⭐ Close to kernels, systems thinking
3 Systems ⭐⭐⭐ ⭐⭐⭐ Good foundation, can transition
4 Networking ⭐⭐⭐ ⭐⭐⭐⭐ Systems thinking, less code-heavy

Career Progression Path

Recommended Strategy:

Entry Point: Systems Engineer (6-9 months)
    ↓ Learn GPU topology, CUDA drivers, hardware
Transition: Inference Serving (9-12 months)
    ↓ Learn inference engines, optimization basics
Target: Inference Optimization (12-18 months)
    ↓ Deep kernel programming, CuTe, peak performance

Alternative Fast Track:

Entry Point: Applied AI Engineer (3-6 months)
    ↓ Learn AI/ML workflows, Python ecosystem
Transition: Inference Serving (parallel study)
    ↓ Study GPU programming alongside work
Target: Inference Optimization (12-15 months total)

Technical Skills Matrix

Core Skills Across Roles

Skill Inf Opt Inf Serve Systems Net Post Train Agentic Internal
CUDA Kernels ⭐⭐⭐ ⭐⭐ ⭐ - - - -
Profiling ⭐⭐⭐ ⭐⭐ ⭐ ⭐ ⭐ - -
GPU Architecture ⭐⭐⭐ ⭐⭐ ⭐⭐ ⭐⭐ ⭐ - -
Triton ⭐⭐⭐ ⭐⭐ - - ⭐ - -
CuTe/CUTLASS ⭐⭐⭐ ⭐ - - - - -
Python/PyTorch ⭐⭐ ⭐⭐⭐ ⭐ ⭐ ⭐⭐⭐ ⭐⭐⭐ ⭐⭐⭐
Linux Systems ⭐⭐ ⭐⭐ ⭐⭐⭐ ⭐⭐⭐ ⭐⭐ ⭐ ⭐
Networking ⭐ ⭐⭐ ⭐⭐ ⭐⭐⭐ ⭐ - -
RL/Training ⭐ ⭐ - - ⭐⭐⭐ - -
Agents/Frameworks - - - - ⭐ ⭐⭐⭐ ⭐⭐⭐

What "Own the Stack" Means

ai& believes in vertical integration:

┌─────────────────────────────────────┐
│  AI Models & Applications           │ ← Agentic AI roles
├─────────────────────────────────────┤
│  Post-Training & RL                 │ ← Post Training role
├─────────────────────────────────────┤
│  Inference Serving & Optimization   │ ← Inference roles
├─────────────────────────────────────┤
│  Kernel & GPU Optimization          │ ← Inference Optimization
├─────────────────────────────────────┤
│  Networking Fabric                  │ ← Networking role
├─────────────────────────────────────┤
│  Systems & Bare Metal               │ ← Systems role
├─────────────────────────────────────┤
│  Data Centers & Infrastructure      │ ← Systems role
└─────────────────────────────────────┘

Your Target: Deep expertise in Kernel/GPU Optimization layer, with understanding of adjacent layers.

Learning Priorities Based on Roles

Priority 1: Essential for Inference Optimization

  • PMPP textbook (complete)
  • CUDA kernel programming
  • Nsight Compute profiling
  • Triton language
  • CuTe/CUTLASS
  • GPU architecture (Hopper/Ampere)

Priority 2: Important for Multiple Roles

  • Linux systems programming
  • PyTorch internals
  • NCCL and distributed computing
  • Python optimization

Priority 3: Nice to Have

  • Agentic frameworks (LangGraph)
  • RL fundamentals
  • Multimodal model serving
  • Web development (TypeScript already known)

Recommended Application Strategy

Timeline: 12-18 months to readiness

Months 1-6: Foundation

  • Study PMPP, basic CUDA
  • Build kernel portfolio
  • Profile existing AI models
  • Contribute to open-source (Triton, CUTLASS)

Months 6-12: Intermediate

  • Master Triton
  • Implement FlashAttention components
  • Study inference engines (vLLM)
  • Build fused kernel library

Months 12-18: Advanced

  • Deep CuTe expertise
  • Optimize kernels achieving 80%+ peak
  • Profile and optimize real models
  • Create technical blog/portfolio

Application Targets

Phase 1 (Month 6-9): Systems Engineer

  • Easier entry point
  • Learn company stack
  • Transition to optimization

Phase 2 (Month 12-15): Inference Optimization

  • Target role
  • Deep technical work
  • Maximum learning

Related Notes

Last Updated: 2026-03-25

Backlinks