Acerca de
If you are looking for work, hit me with…
Artículos de Juan Pablo
-
Azumo on the news
Azumo on the news
Aragon Research has showcased us for their research paper "Hot Vendors in AI Chabots, 2017"
2 comentarios
Actividad
4 mil seguidores
-
Juan Pablo Lorandi ha compartido estoThis is the final piece in the AI infrastructure series. After writing about GPU ownership, production bottlenecks, multi-cloud GPU trade-offs and the hidden infrastructure tax around open-weight models, one pattern keeps coming up in conversations with peers: The model works. The infra drags. Teams stall because production infrastructure becomes heavy, fragmented and expensive to operate. In this article, I break down why the real “last mile” of AI lives in provisioning, serving, integration and operations - not in model capability. It’s also why we built Valkyrie: • Shorter path from POC to production • Lower operational overhead • Clearer cost control across providers • Less engineering time spent stitching GPUs together If you’re navigating that gap between “it works” and “it runs reliably at scale,” this piece is a must. Find the article link in the first comment. And if you want to see how we approached the infrastructure problem, learn about Valkyrie: https://hubs.la/Q045lzbv0
-
Juan Pablo Lorandi ha compartido estoLast week I wrote about the infrastructure tax around open-weight AI and why we built Valkyrie to remove it. Now I’m seeing a related theme: multi-cloud sounds like a strategy… until you try it with GPUs. With CPUs, abstraction works. With GPUs, hardware, drivers, networking and pricing models diverge fast. I wrote about why traditional multi-cloud thinking breaks down under accelerated compute and what a more practical approach looks like. Valkyrie https://hubs.la/Q045lqVH0 Was built for: - Builder teams shipping POCs and first production jobs. - Scale & enterprise teams running regulated or high-throughput workloads. One REST layer. Multi-provider access. No GPU orchestration overhead. If you’re navigating multi-cloud and GPUs, this may be relevant. Full article here: https://hubs.la/Q045lmT90
-
Juan Pablo Lorandi ha compartido estoHave you seen this pattern? The demo looks great. Everyone’s aligned. “Eight weeks to production” Then six months later, it’s still not live. I’ve heard this same story from peers at other companies more times than I can count. The model usually performs as expected. The friction shows up in the infrastructure layer. That recurring pattern is what led us to build Valkyrie at Azumo: a single REST layer to run open-weight models across providers without stitching together GPUs, queues, storage and auth for every project. I wrote about where open-weight deployments stall and what changes when infrastructure stops being the bottleneck. If you’ve taken a model from demo to production, this is worth a read: https://hubs.la/Q045lh6Q0 If you’re curious about Valkyrie, you can learn more here: https://hubs.la/Q045lkGy0Open-Weight AI in Production: Why Prototypes Fail to ScaleOpen-Weight AI in Production: Why Prototypes Fail to Scale
-
Juan Pablo Lorandi ha compartido estoI recently wrote about GPU ownership and how fuzzy accountability leads to cost and operational friction. Today, I'm covering a similar dynamic I see in AI infrastructure discussions. Someone says, “We need data sovereignty.” And the conversation jumps straight to self-hosting. That shortcut deserves a pause. Data sovereignty can mean data residency, regulatory compliance, contractual obligations, or internal risk posture. Those are different constraints and they don’t all point to the same architectural answer. In this piece, I unpack what data sovereignty actually requires, where self-hosting is justified, and where certified cloud or managed platforms meet the same compliance standard without the operational overhead. Clear definitions matter, especially when infrastructure decisions are expensive and long-lived. Check the full article here: https://hubs.la/Q045l9300Data Sovereignty Alone Isn't Worth Self-Hosting: Here’s WhyData Sovereignty Alone Isn't Worth Self-Hosting: Here’s Why
-
Juan Pablo Lorandi ha compartido estoIf I walked into your organization and asked: Who owns GPU infrastructure?, what would you say? Data science provisions GPUs for training and experimentation. DevOps manages infrastructure but may not specialize in GPU workload optimization. IT tracks cost and compliance, often without full context on why certain GPU resources are running. Everyone interacts with it but there is no clear end-to-end accountability. Over time, same things happen: costs creep up, utilization stays lower than it should be, instances keep running longer than expected and there’s no defined standard for performance or reliability across ML workloads. No one formally owns the full GPU lifecycle. I wrote an article breaking down how this ownership gap develops, why traditional infrastructure models struggle in GPU-heavy environments and the structures I’ve seen work. Full article here: https://hubs.la/Q045l8vj0Enterprise GPU Management: Solving the Ownership ProblemEnterprise GPU Management: Solving the Ownership Problem
-
Juan Pablo Lorandi ha compartido estoIn the last two articles, I wrote about why open-weight deployments stall and why self-hosting costs more than expected. This piece focuses on the pattern behind both. API models ship fast. Teams integrate in days, prove value in weeks, and move forward. The infrastructure complexity is abstracted away, so momentum stays high. Then comes the pivot to open-weight. The model benchmarks well. The plan looks solid. GPU requests go out. Framework decisions start. Integration begins. Six months later, the API system is still running production traffic. The self-hosted system is still stabilizing. This isn’t poor execution. It’s structural friction. API providers compress infrastructure into a single call. Open-weight expands it into provisioning, serving, scaling, tuning, and operational ownership. The timeline gap compounds into opportunity cost. If you’re deciding between API, self-hosted, or hybrid, this breakdown will help frame the tradeoffs more clearly: https://hubs.la/Q043dBMh0
-
Juan Pablo Lorandi ha compartido estoIn my last piece I wrote about why most DevOps teams struggle with open-weight models in production. Now: the cost myth. “We’ll just self-host. It’ll be cheaper.” GPU pricing makes it look obvious. The real cost shows up elsewhere: – 4–6 month rollout – Multiple engineers fully allocated – Ongoing on-call and optimization – Peak overprovisioning – Infrastructure complexity Most teams underestimate total cost by 3–5x. Self-hosting makes sense at real scale. Below that, the math changes fast. Full breakdown here: https://hubs.la/Q043dxWd0
-
Juan Pablo Lorandi ha compartido estoRunning open-weight models in production requires more than spinning up GPUs. Llama 3 or Mixtral in a test environment proves very little. Production brings: – Hard GPU memory limits – CUDA and driver compatibility risk – Network bottlenecks in distributed inference – Model loading and storage constraints – Ongoing operational overhead But DevOps teams discover the gap weeks after deployment. This article breaks down the actual constraints and how to evaluate readiness before committing to self-hosted inference. Full piece: https://hubs.la/Q043dvkW0
-
Juan Pablo Lorandi ha recomendado estoJuan Pablo Lorandi ha recomendado estoI am announcing a new game today -- but first, a short backstory :) As you may know, maybe 18 months ago I started a teensy indie game studio called Spyre. I got a couple publishing deals, dropped a few games (Hangman Clash + The Fallen Court -- iykyk), hit over a quarter million skill-based PVP matches (pretty cool), and then continued my irresponsibility streak and decided to pay it forward. Like an idiot, I put some of the revenue into someone else's indie game. Bonkers, I know-- but I've been following along with the game's progress for months and daggum: it's really shaping up! I'm genuinely excited to share this game that I didn't make! Well today is the day to announce this mysterious game. It's called "Flame, Forest & Flood", and it's a three-player, skill-based strategy game crossed with a cozy sim. Doesn't that sound amazing? You will be sipping hot chocolate in fuzzy socks while bringing your adorable enemies to your knees. FFF appealed to me because I wrote Hangman Clash with a similar premise: make a skill-based PVP game that is simple to learn, but difficult to master. That combination is rare, and it is hard to build. Congrats to Blazej Zywiczynski and the good folks over at Fairy Mount Games as their hard work over the past few years is now starting to see the light! The Steam page is live now, and the playable demo goes up June 15 for Steam Next Fest. If you've got ten seconds, wishlist it. This would be enormously helpful as Steam's whole visibility machine runs on that number, and for a small team it can really help. https://lnkd.in/gXacZbiS
Experiencia y educación
-
Azumo, LLC
Mira la experiencia completa de Juan Pablo
Mira su cargo, antigüedad y más
o
Proyectos
-
GorbGames
- actualidad
CTO. Lead Dev Game Engine + Networking code.
-
cSuite
Created several UI solutions for cSuite, an Android application. A lot of the UI/UX work was done using OpenGL
-
Games
- actualidad
Developed Slot Machines for casinos in the US.
Developed Sweepstakes games for clients in Canada and Bolivia
Developed Android Casual Games.
Otros creadores -
Idiomas
-
English
Competencia bilingüe o nativa
-
Spanish
Competencia bilingüe o nativa
Recomendaciones recibidas
2 personas han recomendado a Juan Pablo
Unirse para verloPerfiles similares
Ver más publicaciones
-
Simpat Tech
We’re excited to once again support the .NET community, and even more thrilled to launch the first of many Spanish-language technical talks hosted directly on the official .NET channel😎🤟! To kick things off, we’re featuring Jesús Flores, a seasoned developer from San Nicolás de los Garza, Nuevo León, with over 18 years of hands-on experience working with the .NET ecosystem. Jesús is known not just for his deep technical expertise, but also for his relatable, practical approach, and yes, he once got the Mexican tax authority (SAT) to delay the rollout of a major digital document standard (Anexo 20). True story. In this session, Jesús will dive into how to build reliable and maintainable integration tests using Testcontainers in .NET, leveraging Docker to simulate real dependencies like databases, queues, and more. If you’ve ever struggled with flaky tests or inconsistent dev environments, this talk is for you. Whether you're a senior engineer or just getting started, this is a great opportunity to: -Learn battle-tested strategies from a real world expert -Engage in Spanish with fellow devs from across the globe -Be part of a growing community of professionals who share your passion for building better software Let’s learn, share, and grow juntos🙌🚀! Register here: https://lnkd.in/erDDDc4G #DotNet #Testcontainers #IntegrationTesting #TechCommunity #Docker #SpanishTech #ComunidadDev #SoftwareDevelopment #DevOps #EngineeringExcellence
1 comentario -
UXUI Fans
Web Components nativos: La alternativa ligera para componentes UI reutilizables. Tired of the bloat of frameworks when creating UI components? Discover how native Web Components offer a powerful and efficient solution for code reuse in front-end development. Learn how to build robust components with native JavaScript, improving the performance and compatibility of your projects. [Enlace al artículo] #webcomponents #frontend #javascript #uidesign #uxdesign #webdevelopment https://lnkd.in/ejTDwyVn
1 comentario