Scale-Up
Foundations
Analyst
Refinement
Expansion
Development
Building regional intelligence
From classifiers to continuum thinking
Training, testing, and deployment
From tool to AI assistant
Dissemination and global engagement
From keywords to context
Subtítulo
Subtítulo
Subtítulo
Subtítulo
Subtítulo
Subtítulo
Partnerships across Latin America to build datasets grounded in local vocabulary, slang, and cultural context.
Early model development and a conceptual shift: reframing abuse detection from discrete categories to a continuum.
Iterative improvement of models using real data from Te Protejo, culminating in the first international presentation of the work
ARTEMISA evolves from flagging words to understanding meaning, including cultural references, coded language, and offender community dynamics.
A ChatGPT-style interface makes ARTEMISA's analysis accessible to any analyst, not just data scientists.
International presentations and collaborations across four countries — ARTEMISA recognised as a good practice example of AI-driven OCSEA detection in Latin America.
From Colombia to the world By 2025 and 2026, ARTEMISA's reach had extended well beyond its origins. The team was presenting at major international conferences, collaborating with partners across four countries, and being recognised as a good practice example of AI-driven OCSEA detection in Latin America. Milestones: 1. March–June 2025 — Joint collaboration with partners in Argentina, Costa Rica, and Peru to build shared regional dictionaries 2. April–May 2025 — Acceptance of ARTEMISA presentations at Stanford Trust & Safety Conference and ISPCAN
Making the tool work for analysts, not just data scientists The team developed a conversational interface, allowing analysts to interact with ARTEMISA directly, ask questions, and receive summaries in plain language. The goal was accessibility: powerful analysis shouldn't require technical expertise to use. Milestones: 1. September 2024 — Participation in the Research Convening of the Oxford Internet Institute and World Childhood Foundation 2. November 2024 — Pilot of ARTEMISA as a GPT-style tool with WebIQ, the platform managing the dark-web data 3. March 2025 — Launch of data analysis using the LLAMA-4 model, improving processing efficiency 4. March 2025 — International dissemination of the project through WebIQ's newsletter
The tool needed to speak Latin American Spanish Expanding beyond Colombia meant more than translating the software. The team worked directly with police, attorneys, and local partners across Costa Rica, Peru, and Argentina to gather region-specific vocabulary, slang, emojis, and cultural references, feeding this into ARTEMISA so it could understand abuse patterns as they actually appear across different local contexts. Milestones: 1. November 2023 — Launch of regional project across Colombia, Costa Rica, Peru, and Argentina 2. 2023–2025 — Partnerships with local stakeholders to build linguistically and culturally grounded datasets 3. April–June 2025 — Creation of regional dictionaries to improve model accuracy across Latin American Spanish variants
A new way of seeing abuse Early CSAM detection tools treated abuse as a set of discrete categories. The Los Andes team challenged this, building models that understood abuse as a continuum, where a single case could span multiple stages and types. This conceptual shift became the foundation for everything that followed. Milestones: 1. October 2021 — Project launch: predicting offender-conversation patterns using the Te Protejo hotline and dark-web data 2. November 2021 — Design of first classifiers for Te Protejo hotline reports and dark-web offender conversations 3. January 2022 — Development of first data-annotation system for OCSEA cases in Colombia
Putting the model to work With the conceptual framework in place and funding from the Research Fund, the team focused on making the models operationally reliable through iterating with real data from Te Protejo and refining analytical accuracy through repeated testing and deployment cycles. Milestones: 1. January 2022 – August 2023 — Iterative refinement of analytical models using BERT (Bidirectional Encoder Representations from Transformers, a machine learning framework for natural language processing) 2. August 2023 — Delivery and hands-on training with Te Protejo hotline analysts and the research team analysts. 3. September 2023 — First international presentation at Stanford Trust & Safety Conference
Understanding communities, not just words A keyword-detection tool flags terms. ARTEMISA evolved to understand meaning — recognising implicit references, cultural shorthand, and the social dynamics of offender communities online. This included interpreting song lyrics used as coded language, understanding how exploitative forums function as communities of practice, and reading actions and interactions, not just words. Milestones: 1. February 2024 — Development of second-generation data-annotation system for dark-web offender conversations
Andes
SafeOnline
Created on May 18, 2026
Start designing with a free template
Discover more than 1500 professional designs like these:
View
Project Roadmap Timeline
View
Step-by-Step Timeline: How to Develop an Idea
View
Artificial Intelligence History Timeline
View
Practical Timeline
View
Timeline video mobile
View
History Timeline
View
Education Timeline
Explore all templates
Transcript
Scale-Up
Foundations
Analyst
Refinement
Expansion
Development
Building regional intelligence
From classifiers to continuum thinking
Training, testing, and deployment
From tool to AI assistant
Dissemination and global engagement
From keywords to context
Subtítulo
Subtítulo
Subtítulo
Subtítulo
Subtítulo
Subtítulo
Partnerships across Latin America to build datasets grounded in local vocabulary, slang, and cultural context.
Early model development and a conceptual shift: reframing abuse detection from discrete categories to a continuum.
Iterative improvement of models using real data from Te Protejo, culminating in the first international presentation of the work
ARTEMISA evolves from flagging words to understanding meaning, including cultural references, coded language, and offender community dynamics.
A ChatGPT-style interface makes ARTEMISA's analysis accessible to any analyst, not just data scientists.
International presentations and collaborations across four countries — ARTEMISA recognised as a good practice example of AI-driven OCSEA detection in Latin America.
From Colombia to the world By 2025 and 2026, ARTEMISA's reach had extended well beyond its origins. The team was presenting at major international conferences, collaborating with partners across four countries, and being recognised as a good practice example of AI-driven OCSEA detection in Latin America. Milestones: 1. March–June 2025 — Joint collaboration with partners in Argentina, Costa Rica, and Peru to build shared regional dictionaries 2. April–May 2025 — Acceptance of ARTEMISA presentations at Stanford Trust & Safety Conference and ISPCAN
Making the tool work for analysts, not just data scientists The team developed a conversational interface, allowing analysts to interact with ARTEMISA directly, ask questions, and receive summaries in plain language. The goal was accessibility: powerful analysis shouldn't require technical expertise to use. Milestones: 1. September 2024 — Participation in the Research Convening of the Oxford Internet Institute and World Childhood Foundation 2. November 2024 — Pilot of ARTEMISA as a GPT-style tool with WebIQ, the platform managing the dark-web data 3. March 2025 — Launch of data analysis using the LLAMA-4 model, improving processing efficiency 4. March 2025 — International dissemination of the project through WebIQ's newsletter
The tool needed to speak Latin American Spanish Expanding beyond Colombia meant more than translating the software. The team worked directly with police, attorneys, and local partners across Costa Rica, Peru, and Argentina to gather region-specific vocabulary, slang, emojis, and cultural references, feeding this into ARTEMISA so it could understand abuse patterns as they actually appear across different local contexts. Milestones: 1. November 2023 — Launch of regional project across Colombia, Costa Rica, Peru, and Argentina 2. 2023–2025 — Partnerships with local stakeholders to build linguistically and culturally grounded datasets 3. April–June 2025 — Creation of regional dictionaries to improve model accuracy across Latin American Spanish variants
A new way of seeing abuse Early CSAM detection tools treated abuse as a set of discrete categories. The Los Andes team challenged this, building models that understood abuse as a continuum, where a single case could span multiple stages and types. This conceptual shift became the foundation for everything that followed. Milestones: 1. October 2021 — Project launch: predicting offender-conversation patterns using the Te Protejo hotline and dark-web data 2. November 2021 — Design of first classifiers for Te Protejo hotline reports and dark-web offender conversations 3. January 2022 — Development of first data-annotation system for OCSEA cases in Colombia
Putting the model to work With the conceptual framework in place and funding from the Research Fund, the team focused on making the models operationally reliable through iterating with real data from Te Protejo and refining analytical accuracy through repeated testing and deployment cycles. Milestones: 1. January 2022 – August 2023 — Iterative refinement of analytical models using BERT (Bidirectional Encoder Representations from Transformers, a machine learning framework for natural language processing) 2. August 2023 — Delivery and hands-on training with Te Protejo hotline analysts and the research team analysts. 3. September 2023 — First international presentation at Stanford Trust & Safety Conference
Understanding communities, not just words A keyword-detection tool flags terms. ARTEMISA evolved to understand meaning — recognising implicit references, cultural shorthand, and the social dynamics of offender communities online. This included interpreting song lyrics used as coded language, understanding how exploitative forums function as communities of practice, and reading actions and interactions, not just words. Milestones: 1. February 2024 — Development of second-generation data-annotation system for dark-web offender conversations