I build AI-powered products from the first customer conversation to production.

Founder · Own product · 2026 → now

Arnie — 24/7 AI customer support for small and medium businesses

Designed and built end to end, on my own — interface, backend, AI pipelines and the go-to-market around them.

Arnie answers customers across WhatsApp, voice, and web chat — so businesses that can't staff a round-the-clock team never miss a conversation. Built for local retailers, service businesses, and real-estate agencies across Latin America.

• Shared context engine across chat and voice — PDF and website ingestion, chunking, Postgres full-text + trigram retrieval
• Per-tenant data isolation across client companies
• Agents that execute real actions via function calling, not just respond

TypeScript · NestJS · Next.js · PostgreSQL · OpenAI · Retell AI · ElevenLabs

View the assistant configuration screen at full size

View the dashboard at full size

View the analytics screen at full size

Software Engineer · Highpoint Growth Networks · AI systems · 2024–2026

Giving four AI assistants write access to one production database

AI assistants running in production against real customer data with no unauthorized writes.

Problem

Four different user types — venues, advertisers, admins and internal messaging — needed AI assistance over the same production database, on a platform running across thousands of active screens and hundreds of venues and advertisers. Permissions and context differed radically per role, and a single assistant with direct write access was not acceptable: a hallucinating model could corrupt billing records or leak data across accounts.

Decision

A role-based multi-agent system built on OpenAI function calling, where each role exposes its own scoped toolset. The core pattern is propose/execute: the model never writes to the database — it produces a structured proposal that a human confirms before execution. For admin reporting, a scoped read-only SQL sandbox allows arbitrary queries with zero write surface.

Result

The propose/execute pattern became the team's default for any AI-driven feature touching persistent state.

OpenAI function calling · NestJS · Next.js · PostgreSQL

Software Engineer · Highpoint Growth Networks · Billing · 2024–2026

Moving billing ownership in-house without dropping a subscription

Hundreds of active subscriptions integrated with zero disruption to existing customers.

Problem

Customers were billed through an external processor, but Captiview — Highpoint's ad-tech platform — held no record of it. The client wanted billing to run inside Captiview while keeping the processor underneath. The obstacle was that both systems already held records for the same customers and disagreed: an account Captiview showed as onboarded in August had existed in the processor since March. There was no shared key between them.

Decision

Rather than reconciling by hand or trusting one side, I built a system to do it: pull every customer from the processor, match each one to its Captiview account, and compare them field by field so conflicts surfaced instead of being averaged away. Where the two disagreed, Captiview won — its record was the newer one — and that rule was fixed before the first record moved, not decided case by case. The processor's customer ID was then stored on the Captiview account permanently, so every invoice pulled afterwards attached itself to the right customer.

Result

Captiview became the system of record for every subscription, with the full invoice history attached to the customer it belonged to instead of living in a third-party interface.

NestJS · PostgreSQL · Payment processor API · Data reconciliation

Full Stack Engineer · UseTeam · Voice AI · 2025 → now

Turning resident phone calls into structured, routed records

Deployed across 50+ residential communities in Spain.

Problem

The building owner took the incident calls himself and could not answer them all. Residents reported by phone, in their own words and at any hour, so reports arrived incomplete or never arrived — and nothing kept a consistent record of what had been said or when.

Decision

A voice agent on Retell AI that runs an interview rather than a transcription. It asks what happened, how long ago, and whether the caller is on site, and does not close until it has all three. The answers become a categorised incident written to the system, and the owner is notified with the category already applied. Each call is recorded and scored for tone, so the owner can see whether the person was calm, nervous or angry before deciding what to handle first.

Result

The owner stopped being the phone line. Every report is logged, categorised and delivered with the caller's state attached, without anyone answering or transcribing it by hand.

Retell AI · NestJS · Next.js · Sentiment analysis