I design and build the systems behind customer support for products with millions of users: AI voice and chat agents, intent routing, fraud controls and Voice of Customer loops. 16+ years, four support functions built from zero, currently running support across LATAM.
App in the Air
Aviasales
Not frameworks or audits for their own sake. Each of these exists today, in a live operation, with a number attached to it.
Multilingual voice and chat agents wired into the helpdesk: intent classification, refund and payout flows, fallback rules, and a clean ticket at the end of every conversation. The hard part is never the model, it is the escalation design around it.
Response curves, deflection rate, cost per contact and CSAT by intent, then the three changes that move them. Automation comes last, because automating a broken process only makes it fail faster.
Transaction analysis, real-time alerts, automated blocking and contestation workflows, built in-house instead of renting a third-party anti-fraud vendor that never sees your specific abuse patterns.
Moving the tickets is the easy half. The real work is deciding what the new system should inherit: which history stays operationally useful, which process debt gets left behind, and how reporting survives the move.
Support data, proactive interviews and outbound surveys, including the silent cohorts nobody asks, turned into a prioritised list Engineering can act on. Support belongs at the start of the product cycle, not at the end of it.
Four projects that show how the work usually goes: a constraint arrives, the market offers an expensive answer, and something smaller gets built instead.
A new market could not launch without a phone line, and outsourcing it meant paying vendor margin for a service that would still hand every real case back to the team.
A multilingual AI voice line assembled solo in three weeks: Vapi, Twilio, Deepgram and ElevenLabs for the call layer, GPT for reasoning, n8n for orchestration, Zendesk for the ticket trail. The expensive part was never the model. It was interruption handling, silence, fallback rules and deciding exactly when the bot should stop talking.
App in the AirSupport, QA and community all lived in the same three people, while bookings worth $10M+ a year ran through the product and every incident landed in the same queue.
An automated support system with NLP ticket classification and a self-service knowledge base, real-time SLA dashboards and alerts, support-led pre-release QA, an in-house fraud detection layer, and a five-tier loyalty program. The Freshdesk to Zendesk migration moved 100,000+ tickets and 100+ articles in three months without breaking reporting.
A consumer AI app with hundreds of millions of downloads, where ratings drive installs and every unanswered review is both a support ticket and a public signal.
A review triage operation handling 10,000+ App Store and Google Play reviews a month, a knowledge base written from the actual ticket taxonomy, a structured bug-reporting channel into Development, and refunds automated across Zendesk, Stripe and Paddle.
AviasalesComplaints arrived on social media, email, web and mobile, agencies and airlines had no direct line during incidents, and nothing was measured.
A multichannel support function from scratch for a 10-person team, KPI-driven routing on Zendesk and HelpScout, a self-service help centre, and a crisis-management platform connecting agencies, airlines and customers. Earlier, in marketing, the first affiliate program that later became TravelPayouts.
Hands-on, not vendor-managed. I build the automations myself before asking a team to run them.
I started in social media and community at a travel metasearch company, moved into building its support function, then spent six years running Customer Success and fraud prevention for a travel app with 7M+ users, and founded a music-travel startup along the way. Today I lead support for a LATAM mobility platform across Brazil, Mexico and Chile.
The through-line is the same in every one of those: support is treated as the last layer of a company, and it works far better as the first. The fastest way to change that is to make the operation measurable and then let automation take the volume nobody should be reading by hand.
I write about support operations on LinkedIn, mostly about the decisions that do not fit in a dashboard.
The full architecture, and why the model was never the expensive part.
The technical path existed. The business context said wait, and that was the right call.
First response time got so good that agents started replying before understanding the problem.
Scaling a support function, automating L1, building a voice agent, or untangling fraud and chargebacks: write and describe the shape of it. English, Russian or Portuguese.