Hi, I'm Gregorio.
Product leader, builder at heart. I like staying close to the work: the code, the data, the edge cases everyone else avoids.
These days I'm building LegendCraft, an AI picture-book company I run solo with a small army of AI agents. Six years of work across ML pipelines, real-time ad systems, identity graphs, and consumer AI took me from PM to Director, but the part I actually care about stayed the same: knowing a product well enough to shape what it does and how it gets built.
Off the clock: tennis, squash, golf, trails. Shaped by four cities: Milan, Boston, New York, and now Paris.
LegendCraft
AI-personalized picture books children keep forever.
LegendCraft makes a child the actual hero of a printed, hardcover picture book. Not a face swap stuck on a generic character. The real kid goes into the story: their physical traits, their personality, their fears, the things they're working through in real life. That means a kid in a wheelchair sees a kid in a wheelchair. Same with a birthmark, any hair texture, any background. The hero on the page is a genuinely accurate version of the kid holding the book.
Most AI content is generic and disposable, the kind of thing you generate, glance at, and forget. I think its real power shows up when it makes something permanent instead. LegendCraft is the bridge between the AI world and the physical world: a book a child actually keeps under their pillow.
Under the hood
- Built solo, end to end, and live in production with the whole e-commerce machine running: AI generation, then preview, then Stripe checkout, then print fulfillment, then it shows up on a doorstep.
- A multi-stage generation pipeline: Claude writes the story, age-appropriate and with a moral arc that actually holds together; Gemini renders every page, anchored to composited character reference mosaics. That last part is how I solved multi-character visual consistency, the hardest technical problem in AI picture books and the thing that makes most of them fall apart.
- Every generated page gets scored by a 4-metric evaluation model: identity drift, anatomy, costume, and pose. If a page drifts, I know before the kid does.
- Day to day, the whole thing runs on a multi-agent AI operating system I built myself: PM, engineering, growth, creative, and SEO agents. One person operating as a team.
JobFlow
An open source framework for automating your job applications, built for the agentic era.
JobFlow handles the part of the job hunt that nobody enjoys: everything after you've found a role you actually want. You click a Chrome extension on a posting, and the job lands in a queue. From there it builds a tailored resume from a sealed blueprint, answers the dreaded "tell us about a time" screening questions from your own knowledge base, and tracks every application on a local dashboard from Queued all the way to Interview. Finding jobs was never the hard part. Applying is the grind. This is the thing that does the grind.
Paste your resume into a generic LLM and ask it to tailor itself, and it will happily invent metrics, jobs you never had, and languages you can't speak. JobFlow is local-first and structurally honest. Fabrication isn't discouraged by a polite prompt, it's made impossible by design: sealed blueprints and an ATS whitelist mean the model can only use facts you've already signed off on. Anything outside the lines gets logged as a gap, never quietly slipped into a PDF a recruiter will quiz you on.
Under the hood
- Local-first, and I mean it. A Node server bound to 127.0.0.1, so your career data never leaves your machine.
- Deterministic .docx and .pdf generation with QA gates that check page length, fonts, and margins. The output is the same every time, not a slightly different document depending on the model's mood.
- Five resume blueprint archetypes, with whitelist-gated keyword swaps for getting past the ATS robots without lying to them.
- AGPL-3.0, open core. Fork it, self-host it, take it apart and see how it works. It's my first real open-source attempt, so genuinely, have at it.
Claude Code Foundations
An open-source toolkit for keeping Claude Code sharp as your project grows.
Claude Code Foundations is a toolkit for the one thing that quietly decides whether your AI agent stays sharp or slowly falls apart: its context — what it loads, what it remembers, and how it's structured. You don't have to be an engineer. If you use Claude Code for anything that outlives a few sessions — writing, research, a business, a codebase — it's for you. It's four skills you set up at the start of a project, the highest-leverage hour you'll spend, and they double as a retrofit for a project that's already getting slower and forgetting things.
When Claude Code starts feeling slower and "forgetting" what you told it, the model didn't get dumber, your context did. By default every instruction, note, and skill piles into the agent's working memory and gets re-read on every single message, until the signal you actually need is buried under months of noise. The fix isn't remembering more, it's structure: decide how your project talks to the agent, keep what loads every turn small and high-signal, and pull in depth only when it's actually needed.
Under the hood
- skill-architecture: decide the structure before you build. A thin steering layer, a few "employee" personas with real judgment, narrow doers working under them, and context that loads only when each persona is actually working.
- context-budget: audit and trim what's re-read on every single message. A script measures your always-on tokens and gives you a method to cut them.
- ingestion-compression: route big documents through a subagent so only a short synthesis enters your session, not 20 pages you can never unload.
- permission-setup: per-folder trust you choose, always paired with a hard deny-net for dangerous commands. Prompt-free work without the reckless "skip all permissions" mode, plus a leaked-secret check.
- Proof from a real project: always-on context dropped ~53% (~11,900 to ~5,600 tokens per turn) and the main instruction file went from ~5,800 to ~800 words. A signal-to-noise win, not just a cheaper bill.
Installing substack…
Essays on AI-native building, product leadership, and going zero-to-one solo. Coming soon.
copying files… 61% — estimated time remaining: soon™
Career — click a role to open it
LegendCraft
Founder · 2026 – present
Founded and built LegendCraft alone — an AI-powered personalized picture-book platform, live at legendcraft.io with full e-commerce from AI generation through Stripe checkout to hardcover print fulfillment. The thesis: AI's generative power matters most when it produces something permanent, so children become the genuine heroes of printed stories tailored to their real lives, with representation as a first-class feature.
Architected the whole stack end to end: a React frontend, Supabase backend, and a multi-stage AI pipeline orchestrating Claude for narrative and Gemini for illustration — including solving multi-character visual consistency, the hardest technical problem in AI picture books, plus a 4-metric evaluation model that scores every generated page.
The company runs on a self-built multi-agent AI operating system spanning PM, engineering, growth, and creative agents — one person operating as a multi-person team.
Vibe.co
Senior Product Manager · Oct 2025 – Apr 2026
Owned the identity and data infrastructure layer at Vibe.co, a CTV advertising platform, with a mandate to maximize audience addressability for US advertisers.
Built Vibe's CTV Identity Graph from the ground up in direct collaboration with the CTO, integrating bid-stream, demographic, B2B, and publisher identity signals into a single resolution pipeline — and incubated the Account-Based Marketing business unit from concept to near-GA on top of it.
In Q4 2025 launched the first ABM campaign powered entirely by the proprietary graph, bypassing LiveRamp — proof the graph stood on its own. Also established the company's data catalog, governance, and SOX compliance policies.
Pelico
Lead Product Manager · May 2025 – Oct 2025
Led two modules at Pelico, an AI-driven supply-chain orchestration platform for discrete manufacturers backed by General Catalyst — Microsoft Startup of the Year 2024.
Took a greenfield Inventory Management module from MVP to General Availability within a five-month tenure, onboarding three pilot clients and identifying roughly €1M in inventory-overhang reduction opportunities across them.
Introduced structured agile ceremonies to a team that lacked them, and used AI-assisted prototyping to compress the discovery-to-demo cycle from weeks to days.
Simulmedia
PM → Senior PM → Director of Product · Oct 2022 – May 2025
Joined the TV ad-tech company as a PM and was promoted twice in roughly two years, ultimately directing the data science and analytics roadmap, all data partnerships, and a team of three PMs.
Led end-to-end development of Simulmedia's first ML campaign-optimization algorithm, which improved average campaign performance by 32% and was the central driver of CTV revenue growth from $5M to $30M+ over two years.
Directed the build of a proprietary real-time programmatic bidder from scratch — 50,000 QPS at ≤20ms latency in production — and launched an incrementality measurement product spanning both linear TV and CTV that the sales team used to secure $10M+ in renewals and new revenue within three months.
Along the way: a GenAI data-labeling pipeline that became training data for the reach models, a TransUnion data-monetization partnership worth $1M in net new annual revenue, sign-off authority on the $5M annual data budget, and the MVP launch of Skybeam, Simulmedia's self-serve CTV platform for SMBs.
Madison Logic
Product Manager · Sep 2020 – Oct 2022
Early PM at Madison Logic, a B2B account-based marketing platform, during its scale from $50M to $200M ARR.
Led the scalability overhaul of the internal campaign automation console, cutting campaign setup time by 70%, then exposed the workflow to clients as a self-service portal whose shared-accountability model reduced refunds by roughly 15%.
Built API integrations with Marketo, Eloqua, and Pardot for closed-loop reporting, and won the company's 2021 Technology Impact Award.
Early career
AWorld · Search Fund Accelerator · 2017 – 2020
At AWorld — the sustainability app later selected as the official app of the UN's ActNow campaign — researched and quantified the verified environmental impact of 30+ habits powering the app's impact-tracking engine.
Before that, as a private equity analyst at Search Fund Accelerator in Boston, built financial models for acquisition deal sourcing and increased weekly deal lead generation by 20% while managing a five-person intern team.
Education
Columbia University · Boston University · —
M.S. in Applied Analytics from Columbia University, concentrated in data science and machine learning.
Thesis: an NLP classification model (scikit-learn, NLTK) predicting the political affiliation of tweets with 80% accuracy, using sentiment analysis to map ideological differences across topics.
B.S. in Business Administration from Boston University.
My journey
Off the clock
