Hi, I'm Antoine PedrettiFounder of Helmo Solutions
Cloud/DevOps engineer with scale-up DNA, now building AI products and content in public. Infrastructure, product, and indie hacking — all at once.
I'm a cloud engineer with 3+ years working in the startup ecosystem — shipping production infrastructure at fast-growing companies. One of the platforms I helped scale grew from 0 to 100+ cities in 18 months, processing millions of real-time data points and contributing to a €15M fundraising round.
Working that close to fast scale made me want to be closer to the product itself. So in parallel, I started building my own: AI products, experiments, a content studio, a freelance agency. The empire model — each one teaches me something the next one needs.
I'm also always learning. This year I'm deepening AI engineering, product management, and shipping cadence. I share what I learn as I go.
Active in:
Scale-up DNA, product mindset
Posture stats — the experience without naming employers. Credentials that travel with me, not with the companies I happened to work at.
Projects
Each card is a bet and a lesson. Thesis first, stack second, what I learned third. The story lives on the blog.
DIGEST
AI research digest for arXiv
The daily flood of research is a distribution problem, not a discovery problem. Personalization solves it.
Multi-agent pipelines break in interesting ways at scale. Cache aggressively. Stream everything. Tiered pricing forces architectural decisions.
Stampo
Digital loyalty cards for small businesses
Paper stamp cards work. They also get lost, smudged, and left in the wrong jacket. Loyalty belongs in the phone wallet, next to the boarding pass.
The product is one screen. The platform integrations are most of the build. Apple PassKit and Google Wallet each have their own publishing review and their own way of refusing to talk to you when you skip a step. Pre-launch pentest caught more than the first week of real users did.
verdict-guard
GitHub Action that catches dishonest Playwright tests before they reach main
AI assistants write tests that pass without checking anything — a green toast, a waitForTimeout, no real verification. The Playwright stack reports flakiness but never asks whether the test would actually fail if the feature broke. That reasoning layer is what's missing.
Determinism is the property that makes a score safe to gate merges on. Two layers stay deterministic AST and regex, free and milliseconds. The LLM only runs on real failures and is capped at the entrance to the spend. Cost on a typical PR stays under three cents. A live cross-repo benchmark with predictions recorded before dispatch produced three exact matches.
From the blog
Build stories, scale-up lessons, and indie hacker notes. What I think as I ship.