Technical founder · Agent control, grounding, evaluation

How do you let an AI agent act on real data without trusting it?

I'm Alvaro Zweig. I've built production software for 21 years and co-founded four companies. This year I've worked on one problem: agents that run over real messages and documents, where the model is never the last word.

Read the research See the projects GitHub Email me

01 · Research

Letting AI act on real data without trusting it

This year I've worked on one problem: agents that run over real messages and documents, where the model is never the last word. Three controls I built into Pipe, in production:

AI control

A guard in front of an agent that drafts on its own

Deterministic filter, then an LLM reviewer, then rewrite or bounce, then a human. A separate reviewer model cost a 7-minute VRAM swap per request, so I redesigned the guard around it.

Grounding

No verbatim quote, no claim

Every fact an agent extracts must carry a quote that is checked against the source text. Claims that fail the check are discarded before a person sees them.

Robustness

A tool server that can read and cannot leak

A read-only MCP server written from raw JSON-RPC with no dependencies. It fences tool output against prompt injection, has no outbound egress and logs every call.

Earlier research and talks

Researcher in the AI Group at Universidad de Palermo, Buenos Aires: evolutionary algorithms and machine learning for knowledge extraction.

2012
Customized Opinion Mining using Intelligent AlgorithmsP. Cababie, A. Zweig, G. Barrera, D. López De Luise. CISSE 2010 proceedings, Springer LNEE vol. 151
Read
2011
Automatic Content Extraction on the Web with Intelligent AlgorithmsP. Cababie, A. Zweig, G. Barrera, D. López De Luise. ICAART 2011
Read
2010
Gdarim: classroom allocation with genetic algorithmsOpen-source release by the AI Group, Universidad de Palermo (with P. Cababie, F. Cancelo, G. Barrera)
Read
2023
Guest on “Ecommerce en Latam”, episode 40: Ecommerce in Peru
Watch

02 · Projects

What I'm building

Each page explains the problem, how it works and what it runs on.

Open source · MIT

Pipe

A self-hosted unified inbox with an AI layer you can trust.

WhatsApp, email, Telegram, Teams and more in one place, on your own server. Agents summarize, extract and draft, but every claim needs a verified quote and nothing leaves the box.

How it works

Live

Aldeax

Upload a PDF, practice what you study.

Study communities where your notes and past exams become practice tests. Every question cites the material it came from. Free, no sign-up.

How it works

Open source · MIT

Offline Entertainment

Your own in-flight entertainment.

Save videos and music to your phone from YouTube and ten other sites, then watch them in airplane mode. Runs on your own computer: no accounts, no logs.

How it works

Latest public commits

All activity on GitHub

03 · Gravity

Turn an online store into a marketplace, without replatforming

Gravity is marketplace infrastructure for retailers in Latin America. A retailer plugs it into the store it already has and can sell third-party products in about six weeks. It's built MCP-native, so AI agents can connect to it directly through the Model Context Protocol.

$1MARR

>5×yearly growth

5countries

$0outside capital so far

Recognition

2024
Microsoft Americas Partner of the Year, ISV Innovation, Latin America (winner)Awarded to Hidden in Motion S.A.C., the company that operates Gravity
Source
2025
Professional Achievement Award (Premio Logro Profesional)Universidad de Palermo, School of Engineering
Source

We're raising.

Gravity got here without outside capital. If you invest in B2B software in Latin America, I'd like to talk.

Book a call Write to me meetgravity.io

04 · Career

21 years, four companies founded

2023 – now
Gravity by Quickcomm · Co-Founder & CEO
Marketplace infrastructure for retailers in Latin America. Bootstrapped to $1M ARR across five countries.
2018 – 2023
FedEx Dataworks · Release Train Engineer & Senior Software Architect
Ran the release train across distributed teams and owned the architecture: +10% e-commerce conversion, over $500K a year in savings, and a data-driven prevention engine that cut customer loss by 90%.
2017 – 2019
Decompris.com · Co-Founder
Seed capital from 500 Startups and a grant from the Government of Peru. Launched with 10+ e-commerce integrations.
2017 – 2018
Belatrix Software (now Globant) · Scrum Master & Team Lead
Led teams of 30+ people for US banking and finance clients, and trained industry leaders in agile.
2015 – 2016
Yuppit.pe · Co-Founder & CTO
The first discount search engine in Peru. Selected from 400+ startups by Startup Peru and Start-Up Chile; Seedstars audience award.
2008 – 2015
Hidden in Motion · Co-Founder & CTO
Technology consultancy across Peru, Mexico and Argentina. Clients included BNP Paribas, Belcorp, Finning Caterpillar and Assist-Card.

Full career

2026 – now
Pipe · Founder & Principal Engineer
2023 – 2024
Cámara Peruana de Comercio Electrónico · President, e-commerce marketing chapter
2016 – 2017
Juntoz.com · Development Team Lead
2014 – 2016
Quality Products · Consultant & acting CTO
2013 – 2014
SweatWorks (New Jersey) · Lead developer, Spartan Race e-commerce
2008 – 2013
Cyberjuegos.com · Business Development Manager
2009 – 2011
Universidad de Palermo · Researcher, AI Group
2007 – 2009
Virtuacom · Software Developer
2005 – 2006
e.luma Solutions · Software Developer

05 · About

About me

I started writing software for a living in Buenos Aires in 2005, and spent two years doing AI research at Universidad de Palermo before an MBA at NYU Stern. Since then I've built companies in Peru, Mexico and Argentina, and spent five years on e-commerce architecture at FedEx. This year I built Pipe on top of my own messages, and it changed what I work on: when agents pulled tasks and promises out of real conversations, some of what they reported wasn't there. Now every claim has to carry a quote that checks against the source, and the model never gets the last word. That's the problem I want to keep working on: letting AI act on real data without having to trust it.

Education

MBA, NYU Stern School of Business, 2013
MBA, Universidad de Palermo, 2012
Information Systems Engineer, Universidad de Palermo, 2011

Also

Teaching assistant, object-oriented programming, Universidad de Palermo
ACM ICPC competitor, 2007–2009
SAFe 5 Release Train Engineer · Certified Scrum Master
Spanish and English · Polish (EU) and Argentine citizenship