Statistics with assumptions stated up front, pipelines that keep data trustworthy over time, and systems that answer in milliseconds and stay calibrated as behaviour shifts. Shipping to production is not the finish line: it is where the work starts.
AntioquiaAtlánticoBogotáBoyacáCundinamarcaLa GuajiraMetaValle del Cauca
Applied statistics, data engineering and AI
Modelling, analytics and information systems
Live and in development, with real users
Each one end to end: data model, backend, interface and rollout.
Ministerio del Interior
Policía Nacional
Ministerio de Igualdad
Comisión Legal para la Equidad de la Mujer
Gobernación de Boyacá
Alcaldía de Cali
Gobernación del MagdalenaSystem active
9 services · 3 disciplines
/Data science
Ambiguous questions turned into defensible estimates, with the assumptions and the uncertainty in plain sight.
03 services
For knowing whether a change actually worked, rather than merely coincided with something. The metric and the sample size are fixed before launch, and the result is read by segment as well as on average.
Delivers: Protocol · Analysis · Documented decision
Ask about this serviceSo a survey with a complex sampling design is not read as if it were a simple random sample. Expansion factors, non-response handling, and intervals computed on the actual design.
Delivers: Estimates · Methodology write-up · Reproducible code
Ask about this serviceFor decisions that depend on how much probability there is, not only on which option is likelier. The model ships calibrated, with its calibration curve in view.
Delivers: Model · Calibration metrics · Limits of use
Ask about this service/Machine learning engineering
Models that leave the notebook and run: bounded latency, fresh data, behaviour under watch.
03 services
For taking a model from the notebook to an API that answers within an agreed time. Latency is measured at the 50th and 95th percentiles, which is where the tail lives.
Delivers: API · Container · Service metrics
Ask about this serviceSo the model sees the same data when it trains and when it runs. Ingestion, transformation and versioned features under a single definition.
Delivers: Pipeline · Data contracts · Automated tests
Ask about this serviceFor catching that the model has started to fail before the business does. Drift and performance monitoring, a defined retraining cadence, and control of the bias the model itself introduces.
Delivers: Monitoring dashboard · Alerts · Retraining policy
Ask about this service/AI engineering
LLM systems measured before they are promised: evaluated retrieval, cost and failure modes.
03 services
For answering over your own documentation while knowing how well it retrieves, not just how well it writes. Retrieval and generation are evaluated separately.
Delivers: System · Evaluation set · Baseline
Ask about this serviceSo an agent can query and act on real systems without being able to break anything. Read-only permissions, execution caps, and validation before every action.
Delivers: Agent · Documented guardrails · Trace log
Ask about this serviceFor knowing whether the system improved or merely changed. Labelled cases, a baseline to compare against, a failure taxonomy with percentages, and cost and latency per query.
Delivers: Eval suite · Failure report · Cost and latency
Ask about this service01
Before touching any data: which decision depends on this, which metric would prove it, and which result would refute it. A question that cannot be refuted is not a question.
02
Sources, data contracts and features defined identically in training and inference. Plus an explicit baseline: with nothing to compare against, any result looks good.
03
The evaluation set comes before the model. Failure modes are documented with percentages rather than anecdotes, and the write-up states what these data cannot support.
04
Served behind an API with a stated latency budget, calibrated probabilities, and explicit guardrails on what the system may and may not do.
05
Data drift, performance decay, and the bias the system itself introduces into what it will see tomorrow. From here the process returns to the first phase: the metric that worked stops working.
Data scientist · Bogotá, Colombia
Six years working with data across two worlds that rarely meet: companies such as Aeroméxico, Farmatodo and HDI Seguros, and public institutions in Colombia. In the first, the hard part is scale. In the second, that the data is sensitive and mistakes are costly.
I design the models, build the pipelines that feed them, and keep them scoring in real time. Once an automated decision depends on the value of a probability, calibration stops being a methodological detail.
Data science
Inference, experimentation and predictive modelling
Machine learning engineering
Served models, pipelines and monitoring in production
AI engineering
LLM systems, retrieval and evaluation suites
Systems development
End-to-end platforms, from schema to production
The tools I use to build models, pipelines and systems that run.
Next.js
Web apps and platforms
Python
Statistical analysis and modelling
R
Statistical analysis and visualization
PostgreSQL
Relational databases
Neon
Serverless Postgres
AWS
Cloud infrastructure
Google Cloud
Cloud infrastructure
Cloudflare
Network, DNS and security
Docker
Containers and deployment
Claude / OpenAI
Generative AI and LLMs
LangChain
Agent orchestration
Databricks
Data and ML platform
Power BI
Visualization and dashboards
GIS / QGIS
Geospatial analysis
Git / GitHub
Version control
Cursor
AI-assisted development
Auth.js
Authentication and security
Vercel
Deployment and CI/CD
Supabase
Cloud backend
What needs building, what evidence backs it, and when it goes live.
A first conversation to understand the problem, at no cost and with no commitment.
Message on WhatsAppConfidential
Project information and context are handled with absolute discretion.
No commitment
I review the case and say honestly whether I can add value or not.