Skip to content
Valentina Vanegas Aristizábal

Economist and AI leader.

I learn fast,1 build across fields,2 and take ideas to production.3

Head of AI & Technology at a startup that builds software for architecture, engineering and insurance firms in the US.

  1. 1 · Learn fast

    Three weeks after I joined, the AI voice agent I designed was on real sales calls.

    July 2026 · AI agents in production
  2. 2 · Across fields

    Agents in production, a team of three developers and six products, a public-sector pipeline over ~25M records, and an open benchmark.

    Four projects below
  3. 3 · To production

    Three months in, I led the production launch of an MCP server that connects one of our products to Claude, ChatGPT and Grok.

    October 2026 · AI agents in production

Selected work

  1. AI agents · At work · Since July 2026

    AI agents in production

    An MCP server that connects a SaaS product for US architecture and engineering firms to Claude, ChatGPT and Grok, and AI voice agents on real sales calls.

    In production

    Read the project: AI agents in production
    Three tiers of agent actions by blast radius. The outermost tier, actions that leave the firm, waits for a person’s approval. Each firm has a switch, and actions can be undone for 7 days. Leaves the firm: a person approves Tier 3 Tier 2 Tier 1 agent action per-firm switch · 7-day undo
    Fig. 1 Agent actions by blast radius. Anything that leaves the firm, like an email or a text to a client, waits for a person’s approval in the app.
  2. Research · Open source · v1.0, the instrument

    cielo-bench

    Check or agree? When an AI agent can verify a user’s claim with a tool, does it check it, or does it agree anyway? An open benchmark with deterministic scoring and no LLM judge.

    Built and validated with simulated agents · Real-model results pending

    Read the project: cielo-bench
    Eighteen cells: three conditions (no tool, tool, source fails) by two premises (true, false) by three pressure levels. The false-premise cells with the tool are marked. true false no tool tool source fails NAAuNAAu 3 × 2 × 3 = 18 cells pressure: neutral, assertive, authority (N, A, Au)
    Fig. 2 The design. Agree-after-check is measured in the marked row: a false premise, after the agent’s own tool returned the contrary evidence.
  3. Data and economics · Public sector · 2026

    Sisbén IV data pipeline

    The first Power BI dashboard of Antioquia's Planning Department and the Python pipeline behind it, built on ~25M records of Sisbén, the survey Colombia uses to target social programs.

    Covers all 125 municipalities

    Read the project: Sisbén IV data pipeline
    Pipeline: six annual snapshots, 2021 to 2026; a Python ETL with DuckDB and PyArrow; about 25 million records with 24 derived columns and quality audits; a Power BI semantic model with 19 tables; a dashboard by municipality for all 125 municipalities. 6 annual snapshots 2021–2026 Python ETL DuckDB · PyArrow ~25M records 24 derived columns · quality audits Power BI semantic model 19 tables · TMDL, DAX Dashboard by municipality all 125 municipalities
    Fig. 3 From six raw annual snapshots to a dashboard by municipality, with the data audited before anyone reads a chart.
  4. Team and product · At work · Since July 2026

    A team and its products

    Three developers across six products, from the first conversation with a customer to the production release.

    Reporting to the CEO

    Read the project: A team and its products
    3developers
    6products
    Fig. 4 Architecture and engineering, P&C insurance, HR and PEO, construction and B2B data.

Talks and youth leadership

Communication is the front I’ve practiced the longest. I’ve competed in debate and public speaking, coached public-school students for Model United Nations, run workshops for young people, and moderated panels on economics and research.

  • Hablemos del BANREP: autonomía, política y el desafío de la tasa de interés

    Moderator · Economics Student Organization

  • 2025
    Jóvenes, Ideas que Transforman

    Co-founder and workshop lead · Global Shapers Community (World Economic Forum)

  • Apr – Sep 2025
    Model UN Debate Coach

    Coach · Universidad Pontificia Bolivariana, for public schools in Medellín

All talks and awards

Now

  • Leading AI and technology at a startup: agents in production, six products and a team of three developers.
  • Building cielo-bench, an open benchmark for AI agents. It’s in progress, and the results will be published here.
  • Writing short notes on what I learn along the way.

This is a now page.

Updated

Contact

Write to me about AI agents, building products with a small team, data and economics, or a talk.

Contact