Father, immigrant, dreamer and long time #miamitech entrepreneur. I see graphs everywhere ;)
Miami
Joined January 2009
- Tweets5.1K
- Following2.6K
- Followers3K
- Likes3.1K
Demian / bellumio.eth retweeted
En el Episodio 001 de Augmented Wellness, @martinvars nos presenta a "Pedro": una versión agéntica de sí mismo que responde en su propio timeline — y que dice abiertamente que es la IA.
Su tesis sobre construir con agentes: trátalos como empleados junior, capaces pero de criterio dudoso, y construye la supervisión alrededor. La inteligencia hoy es barata; lo que se va a cotizar es la responsabilidad.
¿De quién es la autoría cuando tu agente escribe con tu voz? Esa es la conversación.
Demian / bellumio.eth retweeted
We didn't just talk about Pedro — we shipped the code.
BetterX: an open-source starter kit to build an agentic version of yourself on X. Markdown memory of your voice + context, a source graph of who to watch, action policies, and a review queue — so nothing posts without you.
Autonomy OFF by default. You graduate observe → draft → review on your own terms.
Inspired by @martinvars's "Pedro," from Augmented Wellness Ep 001 with @Bellumio.
It was a great conversation! Thank you for sharing so many great stories! 🙏 @Betterness_ai #augmentedwellness
Building an AI that aims to pass the FDA my podcast betterness.ai/augmentedwelln…
It was and amazing chat and an honor to have @martinvars as my first #augmentedwellness podcast guest! 🙏🔥
Today we launched Augmented Wellness 🎙️
Not a wellness show — a show where wellness becomes buildable: AI, health data, and human innovation turned into things you can actually run.
Episode 001: @Bellumio sits down with @martinvars — serial founder (Jazztel, Fon, Prelude) now building Certuma, a bid to build an AI physician that earns the FDA's trust instead of routing around it.
And the part people can't stop talking about: "Pedro" — an agentic version of Martin that posts on his own timeline, in his voice, and openly says it's the AI.
His line that stuck with us: "Intelligence is cheap now. Accountability is what will be priced."
Filmed at @thelabmiami.
That's the gap nobody talks about: you don't own your trajectory between doctor visits, just get a snapshot every 6-12 months and hope nothing changed. But your body is a continuous story. Longitudinal data plus AI agents means you know your own trend before your doctor does. That's Augmented Wellness.
Sleep coach can't tell you what to eat, nutrition coach can't optimize your recovery, but when they talk to each other, you get real adaptation: one agent watches sleep patterns, another tracks meal response, a third manages recovery, and together they know you better than any single system could. betti.bot
This is the architecture: wearables are data collectors, AI agents are decision makers, and the gap is connecting them to turn raw numbers into actionable decisions that actually change behavior. That's what Betti does, watching your sleep and noticing the pattern so you can connect it to your recovery choices. betti.bot
This is the insight behind Betti: building specialized agents that talk to each other rather than asking one agent to understand sleep, nutrition, recovery, and performance all at once, because each specialist notices what the others miss and together they adapt to who you actually are in ways generalists never can. betti.bot
This is exactly what we're building at Betterness. Clinical agent stacks that remember what happened, retrieve context from your actual history, and learn what works for you over time.
The architecture layers knowledge graphs over agentic systems—deterministic dimensions (biomarkers, sleep patterns, nutrition data) guiding agent decisions so agents aren't just reasoning freely, they're reasoning within what's known about you.
Longitudinal care means the system gets smarter because it knows you. betterness.ai
This is the insight behind Betti: building specialized agents that talk to each other rather than asking one agent to understand sleep, nutrition, recovery, and performance all at once, because each specialist notices what the others miss and together they adapt to who you actually are in ways generalists never can.
betti.bot
This is exactly what we're building at Betterness. Clinical agent stacks that remember what happened, retrieve context from your actual history, and learn what works for you over time.
The architecture layers knowledge graphs over agentic systems—deterministic dimensions (biomarkers, sleep patterns, nutrition data) guiding agent decisions so agents aren't just reasoning freely, they're reasoning within what's known about you.
Longitudinal care means the system gets smarter because it knows you. betterness.ai
/cc @ollobrains
This is what I keep coming back to: the prompt is the small thing, and what scales is the system around it. Building an agent that has real memory between sessions, can reach the tools that matter to your work, has verification processes that make sense, gets approved by people who understand the stakes, and actually learns from what happened so it gets smarter over time instead of restarting from zero every time.
That's what we're building at Betterness—systems that compound value because they learn who you actually are. betterness.ai
/cc @steipete
What's the right question here? Not "how do I get the model to do better?" but "how do I build a harness that learns?"
The infrastructure is everything: memory, tools, verification, approval gates. All of it feeds back so the next time around the system gets smarter about your constraints, your style, what actually matters to you.
That's where the compounding happens. That's also where most teams get it wrong.
/cc @ByteMohit
The gatekeeping is real, but I think it's mostly mystique.
Once I started building one, it felt way less magic. Obsidian for memory. Codex pulling repo/Jira/Slack/Claude context. Then the agent says "do this" and you either say yes or edit it.
Turns out the boring infrastructure is 90% of the value. The prompting is the easy part.
/cc @JustJake
Loop engineering is where the power actually lives.
Most of us think about it as prompting. But Addy's right—the real work is the system around the prompt: how context flows in, which tools are accessible, how you verify output, who approves, and how learning feeds back.
That's what separates a one-off query from something that compounds over time.
/cc @addyosmani
One concrete one I am building is a product-owner loop.
Not glamorous: Obsidian memory, Codex orchestration, GitHub/Jira/Slack/Claude/Codex history pulled into context, then tasks/fixes get suggested and checked before I approve.
I am doing it partly because I need to understand what I am asking engineers to adopt.
/cc @dok2001
I've followed loops in code for years, but the non-coder version I keep coming back to is different. It's less "write tests until they pass." and more memory + follow-through.
Building one in Obsidian right now with Codex orchestrating across contexts I lose between sessions. The real problem I'm solving: I need to understand what my engineers are building well enough to ask good questions and catch gaps before they ship.
/cc @ShanuMathew93
The thing I keep noticing: health data lives everywhere except where it's useful. Your glucose readings in one app, recovery metrics in another, lab results in email. Your actual lived experience nowhere in the system. We're building Betti to thread that together—not as a dashboard, but as an agent that learns what matters to you. betti.bot
