Three case studies in operational AI infrastructure. Each system is live, instrumented, and accountable to a real institution — not a slide, not a sandbox, not a proof of concept that never shipped.
The brief. Public mentoring programmes in Central America can reach perhaps a few hundred entrepreneurs per cohort, in person, during business hours. The unmet demand was three orders of magnitude larger — and concentrated outside business hours, in the second shift after the children were asleep.
What we built. Mia is a WhatsApp-native mentor with long-horizon memory, multimodal intake (voice, image, receipts), and grounded answers. Every response cites verified curriculum supplied by the partner institution.
Operational discipline. Continuous evaluation against a human-rated benchmark, an escalation channel to live mentors, and an honest dashboard for the programme directors — accountable to the people who run it.
Clinical operations. Patient-flow scheduling, treatment-plan tracking, billing reconciliation, and patient communication — for a network of orthodontic clinics across El Salvador. Built bilingually; integrated with national billing rails.
Energy operations. Inventory management, route optimisation, and field-team coordination for an energy operator — surfaced where the dispatcher actually works, not in a separate BI tool.
Engineered to fit. Each system is sized to the institution — small enough to be auditable, large enough to be the system of record. Documentation, observability, and a transition plan come with the deployment.