Two engineers talking through a whiteboard diagram of a data pipeline in a Mexico City office, laptops open on the table

/Delivery models · Nearshore AI

Nearshore AI engineers in your working day.

Nearshore AI engineers build LLM features, retrieval pipelines, agents and the evals that keep them honest, from Mexico or Colombia, inside your working day. That matters because AI work moves in short loops. You try a prompt or a retrieval change, read the eval results, talk to product, and try again. This guide covers which AI work suits the model, how to run it, how to keep keys and data safe, and when offshore fits better.

What an AI engineer does on your team

An AI engineer turns a model into a working product feature. That means choosing a model, designing prompts and tools, wiring in your data, measuring quality with evals and shipping the result behind your API. It’s software engineering with one extra problem: the output isn’t deterministic, so you can’t test it with unit tests alone.

This is different from general software work, which our nearshore development page covers. It’s also different from data annotation, where people label examples by hand. AI engineers build the systems; annotators produce the labeled data those systems learn from or get scored against.

Nearshore software development Engineering teams for general product and platform work. Data annotation outsourcing Human labeling and review for training and eval data.

The AI work a nearshore team takes on

Most AI engineering falls into a handful of jobs. A nearshore team usually covers several of them, with one senior engineer owning how they fit together.

Prototyping LLM features Extraction, summaries, drafting and classification inside your product.
Retrieval-augmented generation (RAG) Chunking, embeddings, search and the prompts that use the results.
Agents and tool use Letting a model call your APIs, with limits on what it can do alone.
Evals A golden set, scoring rules and regression evals that run before every release.
Model selection and fine-tuning Comparing hosted and open models on your own eval set.
Data pipelines that feed retrieval, fine-tuning and evals from your systems.
MLOps Deployment, monitoring, prompt and model versioning, and rollback.
Cost and latency tuning Caching, smaller models where they pass the evals, and batching.

Which AI work belongs nearshore

Put AI work nearshore when the requirements are still moving. Early prototypes, new agent behaviors and anything your product and design leads are still shaping need daily conversation. A question about tone, edge cases or what counts as a good answer shouldn’t wait overnight.

Eval results drive most decisions, and they’re rarely clear on their own. An engineer who can share a failing case on a call at 2 p.m., agree a fix with product and rerun the evals before the day ends will move faster than one reading feedback a day late.

Well-specified work is different. Expanding an eval suite to a written spec, building a batch pipeline or clearing an integration backlog can run offshore, reviewed by the nearshore leads.

Mexico or Colombia for AI engineering

Mexico’s Ley de los Husos Horarios, published in the official gazette on October 28, 2022, ended daylight saving time for most of the country. Mexico City, Guadalajara and Monterrey stay on UTC−6 all year. Quintana Roo uses UTC−5, Sonora UTC−7, and Baja California and some border municipalities still change clocks with the US.

In US winter, Mexico City matches Central time. In US summer, it matches Mountain time. So the team is within two hours of every mainland US time zone all year, but that isn’t a full overlap with East or West Coast leads.

Colombia stays on UTC−5 all year, matching US Eastern in winter and Central in summer. That suits companies whose product leads sit on the East Coast. We pick the country per team: the AI and data skills you need first, then the hours your leads keep.

Sources: Mexico’s time-zone law (Ley de los Husos Horarios), Cámara de Diputados Colombia legal time, Instituto Nacional de Metrología

Onshore vs nearshore vs offshore for AI work

A general comparison for AI engineering. Individual partners vary, so use it to frame your questions.

	Onshore	Nearshore (Mexico, Colombia)	Offshore (India, Philippines, Egypt)
Relative cost	Highest	Middle	Lowest
Overlap with US hours	Full	Most or all of the day	Short, or none without shifted hours
Eval feedback loop	Same day	Same day	Next day, through written handoffs
Best fit	Roles that need US residency or on-site access	Prototypes, RAG and agent design, ambiguous requirements	Eval suite expansion, data pipelines, batch jobs, integration backlogs

Onshore vs nearshore vs offshore The full comparison across every service line. Offshore AI engineers When India, the Philippines or Egypt is the better fit.

Roles on an AI engineering team

Titles vary between companies, so agree on the work, not the label. This is how we split it.

	AI engineer	ML engineer	Data engineer	MLOps engineer
Main job	LLM features, RAG, agents, evals	Training, fine-tuning, model selection	Pipelines that feed retrieval and training	Deployment, monitoring, versioning
You need one when	You’re building on hosted or open models	Off-the-shelf models don’t pass your evals	Your data lives in many systems	Models and prompts ship often

Evals are the shared language

An AI team without evals argues about examples. A team with evals argues about numbers it agreed on in advance. Set this up in the first weeks, before the feature grows.

A golden set of real inputs with expected outputs or scoring rules, owned by your product lead.
Regression evals that run in CI on every prompt, model or retrieval change.
A short written note for each eval run: what changed, what improved and what got worse.
Cost and latency tracked next to quality, so a better answer isn’t ten times slower.
A weekly review of failing cases with product and design, live on a call.

Keys, data and security

AI work touches the most sensitive parts of your stack: your data, your prompts and your model provider bills. Treat it with the same controls as production engineering, plus a few that are specific to models.

NIST’s AI Risk Management Framework (AI RMF 1.0, January 2023) and the OWASP Top 10 for LLM Applications are good shared references. OWASP lists risks such as prompt injection, sensitive information disclosure and excessive agency, which your team should test for by name.

Sources: NIST AI Risk Management Framework OWASP Top 10 for LLM Applications

Work runs in your cloud accounts with your API keys, never a vendor’s or an engineer’s personal keys.
Least-privilege roles through your SSO, removed the day someone leaves.
Secrets in your secrets manager, never in code, prompts or notebooks.
Written rules for what data may appear in prompts, eval sets and training data.
No customer data in third-party tools without your written approval.
Code, prompts, eval sets and model artifacts assigned to your company in the contract.

The first two weeks

A good start produces evidence, not slides. By the end of week two you should see a working slice and an eval that measures it.

Week one SSO accounts, repository access, cloud roles and model API access through your keys.
Week one A walk-through of the feature, the data it uses and what a good answer looks like.
A first golden set, even a small one, agreed with your product lead.
Week two A thin prototype running against that set, with results written up.
End of week two A short note on data gaps, risks and what they’d build next.

Choosing a nearshore AI partner

Teams are delivered through partner firms that OTRO selects and manages. You get one contract and one point of contact, with senior engineering oversight on top. Ask any partner to show you an eval report from past work, with client details removed, and to explain how they handle keys and data.

Our own AI work includes LLM-based data extraction for a healthcare platform and a design assistant for modular construction. The case studies show how those were built.

Software development teams in Mexico Dedicated engineering teams for US companies. AI integration services Help deciding what to build and putting models to work. Samacare: AI-Powered Healthcare Platform Optimization LLM-based data extraction in a healthcare workflow. Annex: AI-Powered Design Optimization for Modular Construction An AI design assistant for modular building fabricators.

Frequently asked questions

What does an AI engineer do that a software engineer doesn’t?

An AI engineer works with model output that changes from run to run. On top of normal engineering, they design prompts and tools, build retrieval over your data, choose and sometimes fine-tune models, and write evals that score quality on real examples. Most good AI engineers started as software engineers, so ask for both skill sets in the interview.

Is Mexico or Colombia better for nearshore AI engineers?

Both work, and the choice usually comes down to skills and hours. Mexico has the larger engineering market and stays within two hours of every mainland US time zone all year. Colombia matches US Eastern time in winter and Central in summer, which suits East Coast product leads. We pick the country per team, based on the AI and data skills it needs first.

How much does a nearshore AI engineering team cost?

It costs less than an onshore team and more than an offshore one. The exact figure depends on seniority, the mix of AI, ML and data roles, and team size, so we don’t publish rates. Remember the second bill too: model API and compute spend. Tell us the roles you need and we’ll come back with a written plan.

Who owns the code, prompts and models?

You do. The contract assigns code, prompts, eval sets, fine-tuned weights and other model artifacts to your company. Work lives in your repositories and cloud accounts from the first commit, so nothing sits on a vendor’s systems. Ask your counsel to confirm the assignment covers the partner firm’s engineers too, not only OTRO.

Can a nearshore team work with our customer data?

Yes, under rules you write before they start. Decide which data may go into prompts, eval sets and training data, and mask or remove personal data where the task allows. Keep everything in your own accounts, and approve any third-party tool in writing first. Your counsel should check what your customer contracts and privacy laws allow.

How is AI engineering different from data annotation?

Data annotation is people labeling examples: tagging text, drawing boxes, rating answers. AI engineering is building the system that uses those labels: the pipeline, the model, the retrieval and the evals. Many AI projects need both. Our data annotation pages cover the labeling side, and the two teams can share one eval set.

Plan your nearshore AI team

Tell us what you’re building with AI, the data it touches and the hours your leads keep. We’ll come back with a written plan: roles, a location for each and how we’d run the first two weeks.

Plan your team