We source, vet, and manage hiring so you can meet qualified candidates in days, not months. Strong English, U.S. time zone overlap, and compliant hiring built in.
.svg){kind=logo}
Triton is OpenAI's language for writing GPU kernels without CUDA. If you're building high-performance machine learning systems, Triton lets your engineers optimize GPU utilization while staying in Python. LatAm ML engineers increasingly reach for Triton to build efficient deep learning models and LLM inference systems. South connects you with senior Triton specialists from Brazil and Colombia who've optimized transformer models and inference pipelines at scale. Get matched in 48 hours. Start your search with South today.
Triton is an open-source language and compiler for writing GPU kernels. Developed by OpenAI, Triton bridges the gap between high-level Python and low-level CUDA. Instead of hand-coding CUDA C++, you write Triton kernels in Python-like syntax, and the Triton compiler generates optimized GPU code automatically.
Triton is used for custom GPU operations in deep learning: fused attention kernels, quantization operators, sparse matrix operations, and custom loss functions. Popular models like LLaMA, Mistral, and Falcon use Triton kernels in their training and inference pipelines. The language abstracts away memory management, occupancy calculations, and other CUDA details that make GPU programming tedious.
GitHub shows 10,000+ stars and active adoption by AI labs at Meta, Hugging Face, and Anthropic. The Python ecosystem (vLLM, Flash Attention, Megatron-LM) relies heavily on Triton for high-performance custom kernels. Triton adoption in LatAm is growing, particularly in Brazil where research labs and AI startups are pushing the boundaries of efficient LLM inference.
Triton is not a general-purpose language. It's purpose-built for GPU kernels. If your task doesn't involve GPUs or custom kernel optimization, Triton is the wrong tool. But if you're doing anything performance-sensitive on GPUs, Triton is increasingly the right choice over CUDA.
Hire a Triton expert when you're optimizing GPU-based machine learning systems and need custom kernels. Common scenarios: you're running inference servers and need to squeeze every bit of throughput out of your GPUs. You're training large models and need fused operations to reduce memory bandwidth. You're building custom layers that don't exist in off-the-shelf frameworks.
Don't hire for Triton if you're not working on GPU code. If you're doing CPU machine learning, inference via standard frameworks (PyTorch, TensorFlow), or data processing, Triton is overkill. The sweet spot for Triton is performance-sensitive AI systems where off-the-shelf kernels don't cut it.
Ideal team structure: one senior Triton engineer (5+ years with GPU optimization), one junior ML engineer who writes Python and calls Triton kernels, and one DevOps engineer managing GPU infrastructure. For smaller teams, a senior Triton engineer can handle both kernel development and system design for the first 6 months.
Triton shines in AI companies, inference platforms, and research labs. If you're building a traditional software application with a machine learning component, you probably don't need Triton.
A strong Triton engineer understands GPU architecture, memory hierarchies, and kernel optimization. They know how to diagnose memory bandwidth bottlenecks, optimize for GPU cache utilization, and measure kernel performance via profiling. They've shipped Triton code to production and debugged performance regressions.
Red flags: engineers who've only used Triton in tutorials. Triton optimizations are subtle; production Triton requires deep understanding of GPU memory models. If they can't explain why a kernel is memory-bound vs compute-bound, move on. Also watch for overconfidence in performance claims. Good Triton engineers measure everything with profilers like Nsys.
Junior (1-2 years): Can write basic Triton kernels with guidance. Understands memory layout and some optimization patterns. Needs mentorship on GPU architecture and profiling. Can implement fused operations for existing architectures.
Mid-level (3-5 years): Comfortable designing Triton kernels for novel operations. Understands memory bandwidth bottlenecks and can optimize via tiling, blocking, and loop fusion. Can integrate Triton with PyTorch and TensorFlow. Knows profiling tools and how to identify regressions.
Senior (5+ years): Has optimized Triton kernels for production ML systems. Understands GPU architecture at a deep level (Ampere, Hopper, architectures). Can design kernel fusion strategies and memory-efficient algorithms. Has shipped Triton code in inference servers and training pipelines. Knows when Triton is the right choice and when to use other optimization strategies.
For remote work, Triton engineers in LatAm are typically UTC-3 to UTC-5, giving you 5-8 hours of overlap with US teams. Soft skills: they should be communicative about performance bottlenecks and able to explain GPU constraints to non-ML engineers.
Tell me about a Triton kernel you optimized for production. Listen for: specific metrics (throughput improvement, memory usage), GPU architecture considerations, and how they profiled the optimization. Strong answers mention measuring before and after.
You've written a Triton kernel that's running slower than expected. Walk me through your debugging process. Good answers start with profiling (Nsys, Triton benchmarks), identifying memory vs compute bottlenecks, then trying optimizations (block size, data types, memory layout). They should mention specific tools.
Describe a time you fused multiple operations into a single Triton kernel. Listen for: which operations, memory savings, throughput gains, and challenges encountered. A strong answer mentions understanding when fusion is worth the complexity.
When have you chosen not to use Triton and used a different optimization instead? Maturity signal. A great answer: 'We profiled and found the bottleneck was I/O, not kernel performance, so we optimized data loading instead.' Shows understanding of trade-offs.
How do you stay current with Triton and GPU architecture trends? Strong engineers follow Triton releases, read GPU architecture docs, and engage with the community. They might mention following OpenAI's blog, GPU vendor documentation, or research papers.
Explain Triton's programming model. What's a block? What's a warp? Evaluation: they should know that Triton programs are written in terms of blocks (groups of elements), and Triton compiler generates warps and threads. They should understand that you don't explicitly manage threads in Triton; the compiler does.
You're writing a Triton kernel for matrix multiplication. How would you optimize it for different matrix shapes? Evaluation: they should understand tiling strategies, block sizes, and how to choose these parameters based on matrix dimensions and GPU memory. A strong answer mentions auto-tuning.
What's the difference between global memory, shared memory, and registers in the GPU? How does Triton abstract these? Evaluation: they should understand the GPU memory hierarchy and that Triton abstracts it through tiling operations. They should know that cache management is implicit in Triton.
You need to implement a fused attention kernel in Triton. What's your approach? Evaluation: they should outline the computation (query-key attention, softmax, value multiplication), memory layout considerations (blockwise operations, reductions), and how to avoid redundant loads. Strong answers mention numerical stability.
How would you handle custom data types (like bfloat16 or int8) in Triton? Evaluation: they should know that Triton supports mixed precision operations and that type conversions have performance costs. They should understand when to use lower precision for memory efficiency vs when precision matters.
Write a Triton kernel that computes row-wise softmax for a matrix with shape (B, N). Then optimize it for different block sizes and measure performance.
Scoring: basic softmax correctness (50%), numerical stability (20%), performance optimization (20%), clear documentation (10%). A strong submission includes benchmarks showing throughput vs block size.
LatAm Rates (2026):
US Market Comparison:
Triton engineers are rarer than general ML engineers, so premium salaries apply. LatAm Triton talent is concentrated in Brazil (São Paulo) and Colombia (Bogotá, Medellín), where AI labs and startups are pushing GPU optimization. Rates in Brazil tend to be slightly higher due to concentration of AI talent around companies like Nubank and Instituto Tecnológico Vale.
Brazil and Colombia have emerging GPU optimization communities. São Paulo-based AI startups and research labs are increasingly using Triton for inference optimization. Colombian engineers trained in GPU programming have strong fundamentals and are eager to work on cutting-edge AI infrastructure. The region has been a hub for deep learning research (UNICAMP, USP, UNAM) with growing Triton adoption.
Time zone alignment is valuable for GPU infrastructure work. Most LatAm Triton engineers are UTC-3 to UTC-5, giving you 5-8 hours of real-time collaboration with US teams. For GPU optimization work, synchronous debugging and profiling sessions with your Triton engineer are worth thousands in avoided trial-and-error.
English proficiency is high among LatAm AI engineers. They've learned Triton, PyTorch, and GPU optimization through English-language documentation and research. Communication about complex GPU behavior is clear and direct. Cultural alignment is strong: AI engineers in LatAm are pragmatic and metric-driven, which matches the mindset required for GPU optimization.
Cost efficiency is significant. You're saving 40-60% on a LatAm Triton engineer compared to US rates. For a niche skill like Triton optimization, this ROI is exceptional.
Tell us about your GPU workload: are you optimizing inference, training, or custom operations? We match from our pre-vetted network of 500+ LatAm ML and GPU engineers, filtering for Triton experience, GPU architecture knowledge, and the seniority your project requires. You interview 2-3 candidates in 48 hours. We handle ongoing support: if the engineer isn't working out, we replace them within 7 days at no additional cost. Our 30-day guarantee ensures the right fit or your money back.
South's vetting includes GPU kernel benchmarks, GPU architecture assessments, profiling exercises, and code review of Triton implementations. We verify production experience by asking about GPU utilization metrics, latency improvements, and scaling challenges. This filters out candidates who've only done academic Triton projects.
Once matched, you get a fully integrated engineer on day one with visa sponsorship, equipment, and compliance handled. Start matching with Triton experts today.
Triton is used for writing efficient GPU kernels for machine learning: fused attention, custom quantization, sparse operations, and inference optimization. Any performance-critical GPU operation is a candidate for Triton.
Use Triton if you want faster development and automatic optimization. Use CUDA if you need ultimate control or are optimizing for specific GPU generations. For most new projects, Triton is the better choice.
Triton is best for writing custom GPU kernels. Numba is for JIT-compiling Python to CUDA. CuPy is a NumPy-like GPU array library. They solve different problems. Triton is the right choice for kernel-level optimization.
Senior Triton engineers in LatAm cost $115,000-$155,000/year, roughly 40-50% less than US rates for equivalent talent. Brazil and Colombia are the primary sources.
You'll interview qualified candidates within 48 hours of describing your needs. Most placements finalize within 1-2 weeks. Triton talent is rarer, so we move quickly.
For custom kernel development, hire mid-level or senior (3+ years GPU optimization). For integrating existing Triton kernels, a junior engineer can handle most work.
Yes. South places engineers for both full-time roles and project-based engagements (3-6 months). Rates adjust based on engagement type.
Most LatAm Triton engineers are in Brazil (UTC-3) or Colombia (UTC-5), giving 5-8 hours of overlap with US teams. This is ideal for GPU infrastructure work requiring real-time debugging.
We assess GPU architecture knowledge, ask about production kernel optimizations, and request Triton code samples. We verify GPU profiling skills and review actual throughput measurements from their work.
We offer a 30-day guarantee. If the engineer doesn't meet expectations, we replace them at no additional cost. We solve fit issues via intensive onboarding and clear performance metrics.
Yes. South handles visa sponsorship, payroll, tax compliance, benefits, and equipment. One all-in monthly fee; we manage everything.
Absolutely. We've placed teams of 2-4 GPU engineers on inference optimization and training efficiency projects. We ensure team cohesion and shared technical context.
