Staff Augmentation for AI/ML Teams: The Costa Rica Talent Pool

Yes, you can hire AI/ML engineers from Costa Rica. The pool is smaller than the React or Node pools, but the people who are in it are deeper than most US founders expect. This is the practitioner version of that case: what kind of AI/ML work the CR pool actually supports, what to pay, and how to interview so you do not end up with someone who read the OpenAI docs and called themselves an ML engineer.

What “AI/ML engineer” means in 2026

The title is a tent that covers very different jobs. Before you write a role spec, get specific about which of these you actually need.

Applied ML engineer. Trains, fine-tunes and ships models against a real product use case. Lives in Python, PyTorch or JAX, sklearn for the boring parts, fine-tuning on LoRA or unsloth for LLMs. Owns the feature pipeline, the training loop, and a real eval harness. This is the largest bucket of useful AI hires for a product company.

LLM/RAG engineer. Builds retrieval pipelines, prompt scaffolding, agent loops and tool use. Lives in Python or TypeScript, the Anthropic and OpenAI SDKs, a vector store (Pinecone, Weaviate, or pgvector in Postgres), and an eval/observability layer (Braintrust, Langfuse, Arize). LangChain shows up less than it did two years ago, LlamaIndex still earns its keep for indexing pipelines.

ML ops / platform. Owns training infrastructure, model serving, GPU cost, deploy and rollback, drift detection, data versioning. Often a former SRE who got pulled into ML. Increasingly the bottleneck once a product has a few real models in production.

ML research. Reads NeurIPS, prototypes new architectures, runs novel training. Foundation model work. Honest about it: this is not where Costa Rica is competitive at scale.

The first three buckets are most of what product companies actually need. The fourth is rare. Knowing which one you want is the difference between paying $90/h for the right person and paying $90/h for someone who will spin for six months.

What the Costa Rica pool actually looks like

A few honest observations from sourcing here every month.

The applied ML pool is real and growing. The country has computer science programs at TEC, UCR and Cenfotec, and a lot of senior backend engineers have rotated into applied ML in the last three years through internal moves at Intel, Amazon and the local fintechs. They came up through software engineering first, which is a feature, not a bug. They can ship.

The RAG and LLM-ops pool is the strongest of the four buckets. Half the AI work being shipped in CR right now is some flavor of “retrieval over our docs, LLM call, eval suite, observability”. The local senior engineers are good at this because it is software engineering with a model in the loop, and they are already software engineers.

ML ops is thinner but real. Most of the ML ops people we place have a strong SRE or DevOps background and grew into model serving. If you need someone who can stand up a vLLM cluster, set up cost guardrails, and write the runbook, you can hire it here.

Foundation model research is where we tell people to look elsewhere. The pool of PhDs doing novel architecture work in Costa Rica is small. CUDA-level kernel optimization is similarly thin. If your roadmap needs someone who is going to publish at ICML, hire in the Bay Area, Toronto, or London and pay accordingly. We will tell you that up front, the same way we did in the Costa Rica decision framework.

Rate bands

These are real CR senior rates we see in mid-2026. Always ranges, never fixed.

• Senior applied ML / RAG / LLM engineer: $60-$110/h. The top of the band is for people with a real production track record on a high-throughput system and a strong eval portfolio.
• Mid applied ML: $40-$65/h. Solid producers, may need senior oversight on architecture decisions.
• Senior ML ops / platform: $70-$110/h. Tracks the senior DevOps band with a model-serving premium.
• ML research: not a CR rate. If you find it, it is a rare individual, priced by reputation.

For context on how this compares to the broader CR rate map, our roles and rates demand map covers it across the rest of engineering, and the pillar on technical staff augmentation covers the model itself.

Common stacks we see and place

Not exhaustive, just what is most common in our placements right now.

• Python with PyTorch, sometimes JAX for serious training, Hugging Face Transformers everywhere
• Fine-tuning with LoRA / QLoRA on unsloth or axolotl
• Anthropic and OpenAI SDKs, with Bedrock and Vertex for enterprise deployments
• Vector stores: pgvector for “we already have Postgres”, Pinecone for managed scale, Weaviate for hybrid search, Qdrant for the self-hosted set
• Orchestration: LangGraph and LlamaIndex are the survivors, LangChain is in maintenance for most of our clients
• Evals and observability: Braintrust, Langfuse, Arize Phoenix, Helicone. Internal eval suites are increasingly the right answer over off-the-shelf
• Serving: vLLM for self-hosted, Modal and Replicate for managed bursts, Triton for the heavier ops setups
• Data: dbt, Dagster, Great Expectations, and Postgres or Snowflake on the warehouse side

If your stack is in this list, we can staff it. If it is wildly off (Mojo for production, novel inference frameworks, exotic accelerators), tell us in the first call and we will be honest about supply.

How to interview without getting fooled

The single most useful filter we run for an applied ML hire is one prompt: “show me an eval suite for a RAG pipeline you shipped”. Not the architecture diagram, not the model choice. The eval suite.

The reason: anybody can wire an OpenAI call to a vector store in a weekend. The people who have actually run a model in production have an eval suite, because they have been bitten by silent regression. They know about golden sets, LLM-as-judge with its known biases, drift detection, failure mode catalogs, and the difference between offline eval and online eval. If your candidate has none of that and the answer is “we just look at the outputs”, they have not shipped.

Three more signals that work:

A take-home that asks them to instrument a tiny RAG pipeline and add three evals against it. Two hours of work, fully revealing. Hand-grade it yourself.

A live debugging session on a broken prompt or a flaky retrieval. Throw them into a Jupyter notebook with a real failure and watch them work. You will see in fifteen minutes whether they reason about what the model is doing or whether they reach for “let me try a different model” as the first move.

A reading question on cost. “Walk me through how you would cut inference cost on this pipeline by 50%.” A real practitioner has views on caching, prompt compression, smaller models for cheap subtasks, batching, and offline distillation. Someone who only consumed the API for six months will not.

We also wrote up the broader what is technical staff augmentation pillar and the popular roles map if you want the cross-role context. And if your AI use case is wrapped in a larger product build, building custom SaaS covers how we think about scoping that.

When CR is not the right answer

Two cases where we will tell you to staff somewhere else, or to mix.

If you need three to five PhDs doing foundation model research, hire in a Tier 1 US or European city. CR is not where that pool is.

If you need 24x7 on-call coverage with deep ML ops, you need a multi-region team. CR for the Americas-business-hours coverage, plus someone in EU or APAC time zones. Trying to run on-call for ML serving with one CR team is a recipe for burnout.

For everything else, especially applied ML, RAG, LLM ops on production systems, and the engineering layer around models, the CR pool can carry the work. We have the bench. The hybrid pattern that works for most of our clients is a CR senior anchoring the role with mid-level support from elsewhere in LATAM, which we cover in the upcoming how 5e Labs delivers.

Send the use case

If you have a real AI/ML use case in flight and want to see who we would put on it, the fastest path is to tell us the use case in a sentence. We will come back with two to three matched profiles you can interview.

WhatsApp us, usually answered within an hour.