How to Become an AI Engineer in Colombia in 2026

Steps Overview

• Prepare prerequisites, tools, and realistic timelines
• Localize your AI-engineer goal to Colombia’s reality
• Get dangerous with code: Python, Git, and small projects
• Master math, data analysis, and SQL foundations
• Build core machine learning with scikit-learn
• Learn deep learning for vision and Spanish NLP
• Implement MLOps, data engineering basics, and deployment
• Build and ship LLM and generative AI systems
• Extend to systems architect and domain specialization
• Adapt the roadmap: part-time versus intensive bootcamp routes
• Verify your skills and showcase production-ready projects
• Troubleshoot common roadblocks and recovery strategies
• Common Questions

Localize your AI-engineer goal to Colombia’s reality

The perfect “AI engineer in 6 months” diagram usually assumes a flat floor: endless time, dollars, and native English. In a Bogotá or Medellín apartment, the reality is different: full-time work, a budget in COP, maybe a decade since your last math class. Localizing your goal means adjusting the manual to your crooked walls instead of blaming yourself for not matching the drawing.

Start from your constraints, not from YouTube

Before choosing a course or bootcamp, map your reality:

1. Write down weekly hours you can actually study without burning out.
2. List your current assets: degree (if any), math level, English level, savings.
3. Define a concrete 2-3 year target: “ML engineer at a nearshore firm in Medellín” or “AI product builder shipping SaaS from Bogotá.”

Warning: if your plan requires 40 study hours on top of a 45-hour job, the missing screw is baked in. Adjust now, not when you crash in Month 3.

Pick a primary learning lane

With constraints in hand, choose one backbone path and let others play a supporting role:

• University-centric: Deep theory via programs like the Master in Artificial Intelligence at Uniandes, described in detail on international program overviews, or AI-focused tracks at UNAL and EAFIT. Best if you can invest 2+ years.
• Bootcamp-centric: Compressed, practical training through providers like Platzi, IA University, or Nucamp’s 16-week Back End, SQL & DevOps with Python (COP 8,496,000) and 25-week Solo AI Tech Entrepreneur bootcamp (COP 15,920,000).
• Self-directed: Stitch SENA, Misión TIC, Coursera, and Kaggle together; cheapest, but demands discipline and good project selection.

Anchor your goal in the Colombian market

Bogotá and Medellín aren’t practice arenas; they’re active AI hubs. Platforms tracking startups list more than 47 AI-focused companies in Colombia, from healthcare to industrial IoT, in addition to giants like Rappi, Mercado Libre, Globant, IBM, and Accenture. You can see this breadth in datasets like the catalog of Colombian AI companies.

Use that reality to backsolve skills: if you dream of working at a Ruta N-backed startup in Medellín, prioritize Python, ML, and MLOps. If your target is an AI-heavy role at a bank or consultancy in Bogotá, lean into statistics, SQL, and business-facing projects. The right goal is the one that matches both your constraints and the problems local employers are paying to solve.

Get dangerous with code: Python, Git, and small projects

The first real gate in this roadmap is simple and brutal: you must get dangerous with code. Several AI engineer roadmaps, like those dissected by data schools such as TripleTen’s machine learning engineer guide, point out that without solid Python and software basics, every later AI step turns into guesswork. Months 1-2 are about turning your laptop into a tool you can actually think with.

What “dangerous with code” looks like by Month 2

*On the part-time path, these hours spread across roughly two months; intensive learners compress them into four to five weeks.

Your minimal command-line toolkit

On day one, aim to comfortably type commands like python -m venv .venv, source .venv/bin/activate (or .venv\Scripts\activate on Windows), pip install numpy, git init, git add ., git commit -m "first commit", and git push origin main. Being fluent with these means you can set up new projects quickly instead of losing an evening to environment errors.

Small projects with Bogotá/Medellín context

By the end of Month 2, you should have at least three tiny but complete projects: a TransMilenio-style bus notifier that reads a JSON feed and prints next departures; a COP ↔ USD/EUR currency converter that calls a public rates API; and a GitHub portfolio repo organizing your scripts. Courses like the introductory Python specializations on Coursera’s programming catalog pair well with these projects, but the non-negotiable part is that each script runs end-to-end on your machine and lives in GitHub with a short README.

Master math, data analysis, and SQL foundations

Once you’re comfortable scripting in Python, you hit the next crooked part of the room: math you barely remember from colegio and real-world data that’s messy, not textbook-clean. Strong AI programs in Colombia, like the Master in Analytical Intelligence for Decision Making at Uniandes, put heavy weight on linear algebra, probability, and data skills for exactly this reason, as outlined in their MIIA program description.

Focus your math effort

You don’t need every theorem, but you do need the core tools ML actually uses:

• Linear algebra: vectors, matrices, dot products, matrix multiplication, and a basic feel for eigenvalues/eigenvectors.
• Calculus: derivatives, partial derivatives, and the intuition behind gradient descent.
• Probability & statistics: random variables, mean/variance, common distributions, conditional probability, Bayes’ rule, and basic hypothesis testing.

Pro tip: tie every new concept to a tiny Python example - compute a gradient numerically, simulate a binomial distribution, or visualize a normal curve.

Level up data analysis and SQL

In parallel, make Python your calculator and microscope for Colombian datasets:

• Use NumPy for vector/matrix operations and Pandas for tables.
• Plot with Matplotlib or Seaborn to see distributions, trends, and outliers.
• Practice SQL until SELECT, JOIN, GROUP BY, and window functions feel natural.

Free national programs like SENA train around 9 million Colombians per year in technical skills, including programming and databases, giving you low-cost ways to practice foundations before paying for advanced AI courses.

Ship EDA and SQL mini-projects

By the end of this phase, you should have Jupyter notebooks that explore Bogotá traffic accidents or air quality data: clean missing values, compute summary stats, and visualize patterns by hour or neighborhood. Pair that with a small SQL project - say, sales data for tiendas de barrio in Bogotá, Medellín, and Cali - where you write queries for top products, monthly revenue, and high-value customers. These become your first “show, not tell” artifacts for your future AI portfolio.

Build core machine learning with scikit-learn

With Python, Git, and basic math in place, this is where you stop just “writing scripts” and start building systems that learn from data. Career guides like the one from StrataScratch on AI engineer paths consistently stress that a strong grasp of classic machine learning is mandatory before deep learning or LLMs.

Core ML concepts to lock in

Over roughly two focused months, aim to understand and implement:

• Problem types: supervised vs unsupervised learning; when to use each.
• Data splits: train/validation/test, cross-validation, and leakage pitfalls.
• Metrics: MAE, MSE, RMSE, R² for regression; accuracy, precision, recall, F1, ROC-AUC for classification.
• Algorithms in scikit-learn: linear and logistic regression, k-NN, decision trees, random forests, gradient boosting, and k-means clustering.
• Model selection: baselines, GridSearchCV, and simple regularization.

Study rhythm and tooling

Use scikit-learn as your main workhorse, wired to Pandas DataFrames. On the part-time path, schedule 10-15 hours per week for about 2 months; intensive learners can compress into 25-35 hours per week over 4-5 weeks. Every week should include reading documentation, implementing at least one new algorithm, and comparing it to a simpler baseline.

Colombia-tuned ML projects

To avoid toy problems, build models around scenarios local employers understand:

• Telco churn prediction: classify which customers of a fictional Colombian telco are likely to leave; compare logistic regression vs random forest using ROC-AUC and F1.
• Credit risk for financial inclusion: predict loan default using demographic and income features, reflecting how fintechs approach underbanked segments.
• Customer segmentation: use k-means to find purchasing clusters for a Mercado Libre-style marketplace with users in Bogotá, Medellín, and Cali.

Habits that make the learning stick

For every project, write a short report in your repo explaining the business question, features, model choice, metrics, and limitations. Always start from a simple baseline model and only then try more complex ensembles. This discipline sets you up for the debugging-heavy reality of AI teams in Colombia’s banks, retailers, and startups.

Learn deep learning for vision and Spanish NLP

At this point, your linear models and random forests can handle tables, but Colombian companies are also betting on images, audio, and mountains of Spanish text. Enterprise leaders interviewed by BNamericas say this is the moment when they expect concrete returns from AI investments, not just experiments, which pushes teams toward deep learning for vision and NLP in banking, retail, and logistics across Bogotá and Medellín; you can see that shift in their analysis of why enterprise AI is different now.

Key deep learning concepts

Your goal over these months is to understand, not just run, deep models:

• Neural networks: layers, activations, forward and backward propagation.
• Optimization: loss functions, gradient descent, and optimizers like SGD and Adam.
• Regularization: dropout, weight decay, and early stopping.
• CNNs: convolutions, pooling, and why they work so well for images.
• Sequence models: RNNs/LSTMs for sequences and Transformers as the modern default for language.

Frameworks and study rhythm

Pick one major framework - PyTorch or TensorFlow/Keras - and go deep. Complement it with Hugging Face Transformers for language models, especially Spanish variants. Guides on AI careers, like Udacity’s overview of skills and tools for AI engineers, explicitly list these libraries as core expectations for modern roles.

On the part-time track, budget 10-15 hours per week for about 2 months; intensive learners should plan for 30-35 hours per week across 4-5 weeks. Each week should include reading theory, re-implementing small networks from scratch, and then using higher-level APIs for more realistic experiments.

Vision and Spanish NLP projects for your portfolio

To make this relevant in Colombia, build projects that hiring managers at Rappi, Mercado Libre, or local fintechs recognize instantly:

• Fruit quality classifier: use transfer learning with a pre-trained CNN to classify Colombian fruits (lulo, guanábana, mango) as “ok” vs “defect,” connecting to agriculture and export use cases.
• Spanish sentiment model: fine-tune a Spanish Transformer on reviews from local e-commerce sites, classifying opinions as positivo/neutral/negativo.
• Delivery-time estimator: model Rappi-style orders with temporal features to predict delivery time, reflecting traffic patterns in Bogotá or Medellín.

Each project should live in a well-documented repo with a clear README, metrics, and a short write-up explaining architecture choices in plain language - precisely the kind of evidence Colombian teams look for when they need someone who can move beyond scikit-learn into production-ready deep learning.

Implement MLOps, data engineering basics, and deployment

So far your models live happily in notebooks; employers in Bogotá and Medellín need them running as reliable services. Teams at Rappi, Globant, Davivienda, IBM and nearshore consultancies don’t hire “model demo” people - they hire engineers who can ship, monitor, and update models in production. Training guides like ONLC’s AI engineer skills overview list DevOps, APIs, and MLOps alongside machine learning as core requirements, not bonuses.

Expose your model through an API

Your first goal is to turn a scikit-learn or deep learning model into a web service:

1. Pick FastAPI (recommended) or Flask.
2. Load your trained model in a /predict endpoint that accepts JSON and returns predictions.
3. Test locally with uvicorn main:app --reload and curl or a REST client.
4. Containerize:
   • Write a Dockerfile that installs dependencies and starts Uvicorn.
   • Build: docker build -t my-model-api .
   • Run: docker run -p 8000:8000 my-model-api

Pro tip: keep configuration (API keys, DB URLs) in environment variables, not hard-coded in your repository.

Wire in CI/CD and monitoring

Next, automate checks and watch the model in the wild:

• Create a GitHub Actions workflow that runs tests on every push, lints with black/flake8, and builds your Docker image on the main branch.
• Deploy to a small cloud instance (AWS, Azure, or GCP) using your container; start with a single VM or managed container service.
• Log every request and prediction (timestamp, features, output) to a database or log service.
• Add basic metrics: latency, error rate, and a daily snapshot of model performance (e.g., accuracy or MAE on recent labeled data).

Warning: an unmonitored model in production will silently decay as data drifts - treat monitoring as part of the feature, not an afterthought.

Cover data engineering basics

In many Colombian companies, AI engineers also help move data. Aim to:

• Reach intermediate SQL: complex joins, subqueries, window functions for time-based analytics.
• Understand when to use NoSQL (e.g., MongoDB) for semi-structured events.
• Differentiate batch vs streaming pipelines and know, at least conceptually, how tools like Spark handle large datasets.

Timebox this phase

On the part-time path, reserve 10-15 hours per week across roughly 2 months to go from “model in notebook” to “Dockerized API with logs.” If you’re on the intensive track, plan for 30-35 hours per week over 4-5 weeks, ideally ending with one fully deployed ML service that you can show to any employer in Bogotá or Medellín as proof you understand production, not just prototypes.

Build and ship LLM and generative AI systems

By the time you reach this stage, “AI engineer” almost automatically means “comfortable with LLMs.” Employers in Colombia are no longer impressed by simple API calls; they expect you to design systems around large language models that solve real problems for banking, logistics, and customer support. Modern guides, like the generative AI roadmap for beginners, highlight that this work is less about training huge models and more about wiring existing LLMs into useful products.

Master the LLM building blocks

The first task is to understand what you’re orchestrating. Over roughly two months (10-15 hours/week part-time or 30-35 intensive), aim to be fluent with:

• LLM basics: high-level intuition for how GPT, LLaMA, and Mistral-style models are trained and used.
• Prompt engineering: system prompts, few-shot examples, and chain-of-thought strategies.
• Retrieval-Augmented Generation (RAG): embeddings, vector databases, and retrieval pipelines.
• Agents and tools: having an LLM call external APIs and perform multi-step workflows.
• Evaluation: automatic checks, human review, and red-teaming for safety.

Turn prompts into working systems

Once you grasp the concepts, connect them to tools: commercial APIs (OpenAI, Anthropic) or open-source models; orchestration frameworks like LangChain or LlamaIndex; and vector databases such as Pinecone or Chroma. The msmgrad roadmap emphasizes a progression from using AI tools daily to building small generators, then integrating them via APIs - precisely the path you’ll follow as you go from “chatting with models” to shipping an internal assistant for a team in Bogotá.

Use structured programs to accelerate

If you want guardrails instead of stitching everything yourself, Nucamp offers two focused options that fit Colombian budgets. The Solo AI Tech Entrepreneur Bootcamp runs for 25 weeks at about COP 15,920,000, centered on building AI-powered products, LLM integration, prompt engineering, AI agents, and SaaS monetization. The AI Essentials for Work program runs 15 weeks at around COP 14,328,000, aimed at professionals who want deep AI-assisted productivity without becoming full-time engineers. Across programs, Nucamp reports an employment rate near 78%, a graduation rate around 75%, and a Trustpilot score of 4.5/5 with roughly 80% five-star reviews, which makes these options unusually strong for their price point in COP.

Ship Colombia-grounded LLM projects

To make this expertise real, build at least one RAG system on Spanish documents: a chatbot that answers questions about Colombian lease clauses, a tourism assistant that plans safe routes in Bogotá and Medellín, or an internal policy assistant for a fictional local company. Design for guardrails - clear refusals, profanity filters, and strict source citation - and document your evaluation process. That combination of LLM literacy, real data, and safety thinking is exactly what distinguishes “used ChatGPT a lot” from “can architect generative AI systems” in Colombian hiring loops.

Extend to systems architect and domain specialization

When you can already ship models and LLM apps, the question shifts from “How do I build this?” to “How do I design the whole system and choose the right problems?” At this level you’re competing with LATAM engineers who, as regional talent reports note, publish at top conferences like NeurIPS, ICML, and CVPR and are hired globally for AI roles, a trend highlighted by firms that specialize in connecting companies with AI talent in Latin America. The next 18-24 months are about joining that tier.

Think in systems, not notebooks

Systems architects design how data, models, APIs, and users fit together at scale. As you level up, practice answering questions like:

• How do we structure services so one failing model doesn’t bring down the entire product?
• Where do we cache LLM outputs, throttle requests, and handle fallbacks when an external API fails?
• How do we track cost per prediction or per conversation and keep it sustainable in COP?
• What’s our plan for A/B testing new models and rolling back quickly if metrics tank?

Spend dedicated time drawing architecture diagrams for your projects: data ingestion, feature store, training pipeline, inference layer, monitoring, and alerting. Then refactor older projects to fit these architectures.

Pick a Colombian domain and go deep

Architects also have a technical moat: domain expertise. Choose one area where Colombia has real, messy problems:

• Logistics: last-mile delivery in Bogotá traffic and Medellín’s hills.
• Fintech: credit, fraud, and inclusion for people with thin or no credit history.
• Energy and utilities: forecasting demand, managing outages, integrating renewables.
• Public services and health: triaging cases, routing inspections, optimizing scarce resources.

National initiatives backed by organizations like CAF, which is supporting Colombia’s mission to develop AI experts, are betting precisely on this mix of AI and sector depth.

Design 3-4 flagship, end-to-end systems

Over Months 13-24, aim to ship a small portfolio of large projects, each covering data → model → deployment → monitoring:

• A route-optimization and demand-forecasting platform for same-day deliveries in Bogotá and Medellín, combining historical orders, traffic patterns, and weather to recommend driver assignments and paths.
• A credit-risk and fraud detection pipeline for a digital lender, ingesting transaction streams, training both classic and deep models, exposing them via APIs, and adding dashboards for risk analysts.
• An energy consumption forecaster for a regional utility, with time-series models, scenario simulations, and a web dashboard for planners to explore “what if” cases.
• A multilingual customer-support assistant that blends RAG, transaction data, and escalation workflows for a fictional Colombian e-commerce platform.

Structure your 18-24 month extension

A practical way to use this phase is:

• 3-4 months: intensive reading on system design, distributed systems basics, and cost-aware architectures.
• 9-12 months: build and iterate on 2-3 flagship projects, each with production-grade deployment and monitoring.
• 3-6 months: refine documentation, add automated tests, present your systems at local meetups in Bogotá/Medellín, and, if you’re inclined, start exploring research papers or a master’s with a focus aligned to your chosen domain.

By the end, your portfolio shouldn’t just say “I know ML and LLMs”; it should read like a set of internal case studies from a Colombian bank, logistics startup, or utility where you were the one quietly redesigning how the whole system works.

Adapt the roadmap: part-time versus intensive bootcamp routes

Once you know the skills you need, the next decision is tempo: do you bend your roadmap around a full-time job in Bogotá or Medellín, or clear your calendar and sprint? Both paths can work if they’re intentional. Analyses of training options, like a comparative review of machine learning bootcamps on DataCamp’s blog, show that outcomes depend more on weekly hours and project depth than on any single brand name.

The part-time route assumes roughly 10-15 hours per week over 12-18 months. It fits if you’re working, studying, or supporting family. A realistic week might be 2 hours each weekday night plus a 4-6 hour weekend block for projects. You can combine free foundations (SENA, Misión TIC, open courses) with paid, targeted programs. For example, Nucamp’s Back End, SQL and DevOps with Python runs 16 weeks at about COP 8,496,000, giving you Python, databases, and DevOps fundamentals while you continue your current job.

The intensive route looks more like a full-time job: about 30-40 hours per week for 6-9 months. This is where structured bootcamps shine. A common sequence for Colombian career changers is to first complete the Back End, SQL & DevOps program, then move into Nucamp’s Solo AI Tech Entrepreneur Bootcamp, a 25-week path focused on LLMs, AI agents, and productization at around COP 15,920,000. With programs ranging from roughly COP 8,496,000 to COP 15,920,000, flexible payment plans, and live community support in cities like Bogotá, Medellín, and Cali, Nucamp positions itself as one of the most affordable structured options in this space. Reported outcomes of about 78% employment, 75% graduation, and a 4.5/5 Trustpilot rating (with nearly 80% five-star reviews) add extra confidence if you’re betting several months of rent on a program.

To choose your lane, be brutally honest about your constraints and risk tolerance:

• If losing your current salary for 6 months would break you, default to part-time and stretch to 12-18 months.
• If you have savings or support and crave fast immersion, an intensive bootcamp sequence can compress the journey and surround you with peers.
• If you’re unsure, start part-time with one structured course; you can always ramp into an intensive path once you’ve tested your capacity and interest.

The “right” adaptation of this roadmap is the one you can stick to in your actual apartment, with your actual budget and responsibilities, not the one that looks best on a YouTube thumbnail.

Verify your skills and showcase production-ready projects

Titles in LinkedIn bios can be generous; production systems and repositories are not. Colombian employers at Rappi, Bancolombia, or nearshore firms care far less about what you call yourself and far more about whether you can design, ship, and maintain AI systems. Experienced engineers interviewed in roadmaps like “How I Would Become an AI Engineer in 2026 If I Had to Start Over” repeatedly emphasize proof over promises.

“Your GitHub repository is your resume.” - Industry AI practitioner, Data Science Collective community

Use a concrete checklist to verify that resume before you send it anywhere. You’re in good shape when:

• Python & software engineering: you can build, test, and refactor non-trivial modules, use virtualenvs, and work with branches and pull requests without thinking.
• Data & math: you can explain vectors, gradients, and basic probability, and produce clean EDA notebooks from raw Colombian datasets.
• Core ML & deep learning: you’ve trained multiple classical models and fine-tuned at least one CNN for images and one Transformer for Spanish text, choosing appropriate metrics.
• MLOps & deployment: you’ve shipped at least one Dockerized model behind an HTTP API, with logs and basic monitoring.
• LLM systems: you’ve built at least one RAG application grounded in Spanish documents (legal, tourism, or internal policies).
• End-to-end thinking: you have 3-5 projects that go from data ingestion to deployment and documentation.

Next, make those projects legible. Each repo should have a clear README (problem, data, architecture diagram, metrics, limitations), a requirements.txt or pyproject.toml, example requests for APIs, and at least smoke tests. Aim for one “flagship” project you’d be proud to walk through in detail for 20-30 minutes.

Finally, stress-test yourself the way Bogotá and Medellín interview loops will:

• Re-implement a past project from scratch in a weekend, including deployment.
• Explain a model’s failures and trade-offs to a non-technical friend in Spanish.
• Simulate an interview: 45 minutes of live coding, 45 of system design for an AI feature (e.g., a basic recommender or support assistant).

If you can do all that from your slightly crooked desk - and your systems still run on Monday morning - you’re not aspiring to be an AI engineer anymore. You’re already working like one.

Troubleshoot common roadblocks and recovery strategies

Even with a good roadmap, there are nights when the desk is still crooked: your model won’t train, English docs feel overwhelming, or you lose a week to binge-watching tutorials. That’s normal. The difference between people who make it into AI roles in Bogotá or Medellín and those who don’t is rarely “talent” - it’s how they handle these stuck moments.

Some roadblocks show up for almost everyone:

• “Tutorial hell” and no projects: you keep consuming courses but never ship anything end-to-end.
• Math anxiety: you avoid linear algebra or probability because it feels like colegio all over again.
• Tool thrashing: every week you switch between PyTorch, TensorFlow, LangChain variants, and five cloud providers.
• Burnout: full-time job + 25 study hours was never sustainable, but you tried anyway.

Recovery strategies work best when they’re small and brutal-honest:

• Pick one stack for 6 months (e.g., Python + scikit-learn + PyTorch + one cloud) and ignore everything else.
• For every 3 hours of tutorials, spend 3 hours building your own mini-project, however ugly.
• Schedule “math sprints”: 30 minutes a day for 6 weeks on just one topic (vectors, then derivatives, then probability).
• Lower your weekly commitment if you’re missing sessions two weeks in a row; consistency beats ambition.

When self-study keeps stalling, borrowing structure can save your roadmap. Bootcamps like Nucamp are designed for exactly this: affordable programs from about COP 8,496,000 to COP 15,920,000, flexible schedules, and local community chapters in Bogotá, Medellín, and Cali. Their reported ~75% graduation rate and ~78% employment outcomes show how much a clear path and accountability matter for career changers juggling jobs and families.

Finally, stop troubleshooting in isolation. Join meetups at Ruta N, university communities, or online forums. Threads where senior engineers share “what actually matters” - like one widely-circulated discussion on becoming an AI engineer in 2026 - can reset your expectations: fewer shiny tools, more shipped systems. Your apartment may stay crooked, but your process doesn’t have to be.

Common Questions