Nearshore AI agent development from Colombia, built for production
Building an AI agent demo takes days. Running one that handles real customers, follows your rules and stays within budget takes a different discipline. I build and operate production AI agents from Bogotá for US companies that want senior work in their time zone.
What I build
Each agent has one job, a defined set of tools and a named owner on your side who answers for the metric. Scope creep is the most common reason agent projects stall, so the first release does one thing well.
- WhatsApp agents that answer leads in seconds, qualify them, collect documents and hand off to a person with the context summarized.
- Voice agents that take calls, answer within approved scripts and transfer live to a human.
- Document review agents that read IDs, statements, certificates and contracts, check validity, cross-check data and flag what's missing.
- Operations agents that watch your pipeline, spot stalled cases and nudge the right owner.
Production, not prototypes: governance
- Business rules live in code below the model: pricing, eligibility, who the agent can contact and what it can promise.
- Least-privilege access: each agent sees only the data its task needs.
- Human approval for sensitive actions until the track record shows the agent can act alone.
- Full logging of every conversation, tool call and human correction, for audit and improvement.
- Platform rules respected by design, including Meta's WhatsApp template and messaging rules.
Evals and cost per case
Every agent ships with a test set of real cases that runs before each change, so a prompt tweak doesn't silently break something that worked last week. When the agent fails in production, the case goes into the test set.
Cost is measured per resolved case, not per month of tooling. Model choice per task, prompt caching and moving deterministic work into code are the main levers. On a mortgage broker copilot, prompt caching cut the cost per conversation 10x, to about USD 0.41 per broker. Every number is labeled as measured or estimated.
Why Colombia
- Same working hours: Bogotá doesn't observe daylight saving time and lines up with US Eastern or Central time all year.
- Bilingual work: English with your team, native Spanish for Latin American customers.
- Direct experience with WhatsApp-heavy, document-heavy operations, which is where most US companies with LatAm customers struggle.
Track record
Today I design and run the AI agents of a VC-backed mortgage platform in Colombia: 5,100+ sessions in production, an operations agent that absorbed the routine layer of a coordinator role (about USD 51K a year) at USD 274–400 a month, a broker copilot that went from zero to production in 18 days, and a mortgage plugin published in the ChatGPT directory. Before that I co-founded and ran DORA, a rental-guarantee fintech with 2,000 tenants and a 0.8% default rate, which was acquired.
FAQ
- Is this a staffing agency?
- No. You work directly with the person who designs and runs the agents. For larger builds I can lead your internal engineers or a vendor, but the architecture, governance and evals stay with me.
- Which models and stack do you use?
- I pick the model per task based on quality and cost, and avoid locking you into one provider. Agents integrate with your existing systems: CRM, databases, WhatsApp Business and telephony.
- How fast can an agent be in production?
- A well-scoped agent can be operating in weeks. The broker copilot I built went from zero to production in 18 days. Most of the time goes into agreeing on the process and the rules, not the code.
