Info
🏫 Microsoft Certified Trainer (MCT) Azure, Data & AI, Databricks
🎓 Master of…
🎓 Master of…
Artikel von George Pirvu
Aktivitäten
3596 Follower:innen
-
George Pirvu hat dies geteilt💡 If you are using Databricks Genie and costs are spiking, probably you are thinking about how many tokens it burned and at which cost. If this is the case, just stop for a moment and read this. It will be helpful. Genie doesn't charge per token. You pay for the SQL warehouse it runs on, as a compound system. Probably you fell overpaying because of: 1️⃣ pointing to uncurate data or raw tables 2️⃣ oversized warehouses 3️⃣ lack of metadata information that Genie can`t feed in the LLM context 4️⃣ long idle timeouts burning DBUs between sessions 🤖 Costs spike when Genie is pointed at the wrong things. You can actually mitigate and control this. What if I tell you, you could actually use Genie for 100 users in production on a Small warehouse for roughly $6–7 per user per month at list price? We have done the calculation in the article. ❔ What does your BI-team do today when a business user needs an answer that isn’t already in a dashboard? Open ticket, assign to a BI engineer, create new report or modify existing report, publish. Just to answer couple of questions. This friction can be resolved. ❔ Will this billing model change in the future? This I don`t know, but that`s how it works today. 📰 Link to the article in the comment section. #Databricks #DataEngineering #CloudCosts #BI #AIEngineering #Genie Adastra
-
George Pirvu hat dies geteilt💡 Insilico Medicine, a Hong Kong-based biotech using generative AI for drug discovery, finds a clinical candidate in 18 months. Traditional discovery takes 5-6 years. They test 60-200 compounds. Legacy methods screen up to 1 million. Their lead drug Rentosertib hit positive Phase 2a results for idiopathic pulmonary fibrosis, targeting a protein nobody had tried before, with AI. Eli Lilly, one of the world's largest pharmaceutical companies, noticed. $2.75B. Full commercial licensing deal across multiple therapeutic areas. For its part, Lilly is no stranger to AI partnerships, having teamed up with Nvidia in October to build what it called the “most powerful” supercomputer in pharma. AI is accelerating drug development, but it remains to be seen whether those accelerated compounds will pass clinical trials at a higher rate than those developed in traditional ways. 💊 German pharma still has the science base, but if we are still waiting for the next steering committee to approve a ticket change request, this is a losing battle. 👉 Link to article in the comment section. #Pharma #DrugDiscovery #AI #GermanPharma #Biotech
-
George Pirvu hat dies geteilt💡 Prompts Are Not Security Controls I'm seeing a dangerous pattern in AI engineering projects: "Let's enforce the guidelines in the prompt." 💥 This is wrong and critical to understand. Prompts are probabilistic. The model CAN and WILL ignore your instructions. No matter how clearly you write your system prompt, it's not a security boundary. It's a suggestion the model will occasionally break. Never use prompts to enforce: 1️⃣ Access control 2️⃣ Data privacy rules 3️⃣ Compliance requirements 4️⃣ Safety constraints Use securtiy layers instead: 1️⃣ Input validation before the model 2️⃣ Output filtering after generation 3️⃣ Schema enforcement for structured outputs 4️⃣ Human approval for critical actions 5️⃣ Defined network perimeters 6️⃣ Monitor and evaluation Prompts are great for tone, style, and preferences. If breaking thre guardrails has consequences, don't trust natural language to prevent it. LLM can`t predict the outcome of its recommendations and actions. #AI #LLMs #Security #MLOps #Databricks #GenAI #AIEngineering #RAG
-
George Pirvu hat dies geteiltWhy is streaming data becoming more and more popular? 💡 Most data is born as a stream (clicks, sensor spikes, transactions). Batch processing is just the "artificial suppression" of that flow, like holding back the flow until we’re ready to deal with it. 🎯 One reason is the increased popularity of real-time analytics. Businesses are demanding up-to-date data for faster, better business decisions, and fresh data is more actionable. 💫 Streaming data is becoming popular also for operational applications. Critical operational applications need real-time data for effective and instantaneous response. In high-volume manufacturing, the goal is to stop a defective part at Step 2 so you don't waste money and time on it through Step 50. In many systems, reacting late is expensive. Streaming removes the artificial waiting step between event and action. 📰 I wrote an article that explains streaming simple, how it works in Databricks, but alsow why it is challenging to implement. Link in the comments. 👇 #databricks #streaming #spark #ai #machinelearning #dataengineering Adastra
-
George Pirvu hat dies geteilt💥 Something we have been working on behind the scenes is finally real. 🇸🇰 The first-ever Databricks User Group in Bratislava is happening on May 12. A big thank you to the Databricks team for trusting us to build this together. You don't just build technology, you invest in the people using it and building a community, step by step. That`s powerful. 💡 Slovakia, Czech Republic, Poland, Romania, Hungary. People here love data and technology. They get excited about it. They go deep, they experiment, they want to have a place to share and learn from each other. They actually care about building things well. That energy is awsome. And it deserves a place to meet. I am happy that Databricks is also about community. 🤝 Thank you, Darius Rahimi, Viktor Bartusek, Stepan Chop, Tomasz Kurzydym for the very good collaboration over the past months. 🤝 Thank you, Marius Maslo, Luboslav Gabal, Simona Laqua-Willigens & Tomas Synek for enabling this initaitve. 🤖 Pridaj sa k Databricks User Group Bratislava a pomôž formovať databricks komunitu na Slovensku. Ak si v okolí, toto je tvoja miestnosť. #databricks #usergroup #adastra #slovakia #bratislava #kosice #databrickscommunity #aiengineering #dataengineering #slovak AdastraGeorge Pirvu hat dies geteiltDear network, Together with Adastra, we're excited to invite you to the first-ever Databricks User Group Bratislava on May 12! 🇸🇰 Here’s what’s coming up: 🧐 Beyond the Diagram: The Concrete Decisions Behind Each Medallion Layer — Peter Karas, Multitude ⏳ Transitioning from legacy ecosystems to Databricks Lakehouse — Luboslav Gabal, Adastra 🍻 Networking, drinks & good food 🔜 and a 3rd presentation to be announced soon Don’t miss this great evening of learning, sharing, and connecting with the data community in Bratislava. 👉 Save your spot now: https://lnkd.in/dMHwVkD6 Viktor Bartusek Stepan Chop Tomasz Kurzydym George Pirvu
-
George Pirvu hat dies geteiltHello everyone, tomorrow I'm going to present how to build the retrieval layer of an AI application on Databricks, hands-on. 🔴 Live tomorrow. Live Lab #61 - Building the Retrieval Layer for Modern AI Apps with Databricks Vector Search 💡 Here's what we're doing in 60 minutes: PDFs → embeddings → Delta Lake → Databricks AI Vector Search → Similarity Search → Reranking → Results Analysis That's the retrieval backbone of a RAG system. Built on Databricks. 📅 Tomorrow, 06:30 PM CEST 📍 SuperDataScience Platform P.S. I've used a lot of platforms in the past. SuperDataScience is cool. SuperDataScience is the one where you actually leave a session knowing how to do something you couldn't do before. They have great course on AI you might want to check it also. If you are interested,fee free to join the community. Link in comments. 👇 #databricks #genai #rag #aiengineering #llm #dataengineering #lakehouse #mosaicai Adastra
-
George Pirvu hat dies geteilt“AI psychosis” or “delusional spiraling” is an emerging phenomenon where AI chatbot users find themselves dangerously confident in outlandish beliefs after extended chatbot conversations. This phenomenon is typically attributed to AI chatbots’ well-documented bias towards validating users claims, a property often called “sycophancy.” 🤖 The Core Problem AI chatbots trained with RL from Human Feedback develop a bias toward agreeing with users called sycophancy. Because users rate agreeable responses are overwhelming, this paper below formally investigates whether sycophancy causes "delusional spiraling": users becoming dangerously confident in false beliefs after extended chatbot use. 💡 Key Findings Even a perfectly rational Bayesian user spirals into false beliefs when the bot is sycophantic. The paper argues that Sycophancy is the driver, not just hallucination. A bot that hallucinates randomly is less dangerous than one that hallucinates strategically to validate the user. 🎯 Right now there is no obvious solution to completly mitigate this Sycophancy. According to the paper, it must be addressed directly at the training level. Awareness campaigns and RAG guardrails alone are also not enough. Paper was signed by MIT CSAIL, University of Washington Seattle, MIT Department of Brain & Cognitive Sciences Conclusions belong to you. 📰 Link to the paper in the comments. #ArtificialIntelligence #MachineLearning #Databricks #AI #GenAI #RLHF
-
George Pirvu hat dies geteilt💥 Something just went live in Slovakia. The Databricks User Group Chapter Bratislava is officially launched. 💡 Slovakia has a technically strong and professional data community. Bratislava, Košice, the talent is there. What's been missing is a dedicated space where Databricks practitioners and customers can sit in the same room and realize we are all in the same boat when it comes to data challenges. This is the first official Databricks User Group chapter in Slovakia. One community, led by practitioners, for practitioners. Every strong community starts small. What to expect from our sessions: 1️⃣ Lakehouse architecture & Delta Lake patterns 2️⃣ AI & ML on the platform 3️⃣ Unity Catalog governance 4️⃣ Scaling data platforms in enterprise environments 5️⃣ Lakebase & Apps usecases 6️⃣ Great conversations about what's hard and how to solve 📌 Join the Databricks User Group Bratislava and help shape the Databricks community if you are around. 🇸🇰 Pridaj sa k Databricks User Group Bratislava a pomôž formovať databricks komunitu na Slovensku. Ak si v okolí, toto je tvoja miestnosť. 🙏 This wouldn't have been possible without the great collaboration of Darius Rahimi from Databricks and the SK Adastra team led by Marius Maslo and Luboslav Gabal. Proud to be building this together. More to come. #Databricks #DataEngineering #Slovakia #Bratislava #Lakehouse #DataCommunity #AI #LLM #GenAI #Kosice #Slovensko #DatabricksKomunita #SlovenskaKomunita Adastra
-
George Pirvu hat dies geteiltYou need to install a critical library for your Databricks workloads on the shared cluster. For example Oracle jdbc connector. You found the right one in the repository but when trying to install it you get: ❌ "JARs and Maven libraries on shared clusters must be on the allowed list." What does this even means? 💡 Here is why the allow list is necessary. Shared clusters are used by entire teams. One bad library can take everyone down. So Databricks says: nothing gets in unless someone with the right privileges explicitly allows it. That's the allow list. Think of it like a building with a guest list at the door. Doesn't matter how legitimate you are. If you're not on the list, you're not getting in. Same with the library. 📽️ I recorded a short video showing exactly how to do it. It might save you some time. Link in the comments. 👇 #Databricks #DataEngineering #DataGovernance #Unitycatalog #Azure #BigData #DataPlatform #CloudComputing #DataManagement #Spark
George Pirvu hat einen Beitrag kommentiert
1 Tag
Completely disconnected prediction from real world and real customer environment.
-
George Pirvu gefällt dasOh yeah, world models are energy-based factor graphs. I even use the common factor graph symbols in my diagram: filled circles for observed variables, hollow circles for latents, rectangles for factors (additive energy terms). But I also add boxes with a rounded side (like a NAND gate) to represent functions (e.g. a neural net). I think this is necessary to indicate which variables are easily computed from others (by forward prop) and which are not (e.g. inferring latents by optimization, variational inference or whatever).George Pirvu gefällt dasThere is a lot of excitement around world models, JEPA, and energy-based approaches to intelligence. From a robotics point of view, this feels very familiar. Factor graphs already give us a structured way to represent dynamics, measurements, and objectives, and optimization already lets us do both estimation and planning. In a new blog post, https://lnkd.in/eh-FSAFR, I argue that if we anchor factor graphs at the present moment, we get a clean picture: - the recent past as a fixed-lag smoothing problem for perception - the near future as an MPC-style planning problem both tied together by the same state, dynamics, and optimization machinery I call this STAG: Sense-Think-Act via factor Graphs. Curious what people think, especially those working on world models, energy-based models, and robot planning.
-
George Pirvu gefällt dasGeorge Pirvu gefällt daswe at Databricks love being with customers. It is always the best part of my job…especially the day today with PUMA Group at their HQ. Thx for having us Govind, Jan, Mihai & team - Sydne-Aline, Oliver, Thomas & I really enjoyed the time! ✨🤘🏻
-
George Pirvu gefällt dasGeorge Pirvu gefällt dasWhat’s the opposite of speed dating? 🙃 We are at Microsoft today, taking part in Cloud & AI Customer Partner Meet-Up. A full day of 1:1 meetings discussing: ✔️Enterprise AI Platform: Scaling AI Safely Across the Organization ✔️AI for Logistics Optimization ✔️Chat with Your Data: AI-powered analytics for business users ✔️Building a Modern Data Platform, and more! 🔥 And the best thing is that we can really go in depth in these discussions, as we have 60 minutes per meeting. 🚀 To help clients get started and move from talk to action quickly, Adastra can secure a range of Microsoft client acceleration funds to support customer projects across both pre-sales and delivery phases. Ping me if this resonates. We’ve got you covered. 🙌 #Microsoft #Adastra #Cloud #Data #AI #AIGovernance Soňa Nováková Emilie Luedecke Petr Bergl Stanislav Kosík Martin Schaefer Zdeněk Soldán ✌︎ Jaroslav Škrabálek Ondrej Melichar Eduard Pfeiler Petr Žikovský Frantisek Raszyk Alzbeta Bartonickova David Frantik Andrea Barešová Karel Florian Ali Adil Alp Eren Kaplan
-
George Pirvu gefällt das𝗗𝗲𝘀𝗸𝗶𝗹𝗹𝗶𝗻𝗴 𝗼𝗱𝗲𝗿 𝗨𝗽𝘀𝗸𝗶𝗹𝗹𝗶𝗻𝗴 𝗱𝘂𝗿𝗰𝗵 𝗞𝗜? Diese Frage stand im Panel der Plattform Lernende Systeme - Germany's AI Platform auf der DMEA - Connecting Digital Health im Raum – insbesondere mit Blick auf die Ausbildung in Medizin und Pflege. 𝐌𝐞𝐢𝐧 𝐏𝐮𝐧𝐤𝐭 𝐚𝐮𝐬 𝐝𝐞𝐫 𝐃𝐢𝐬𝐤𝐮𝐬𝐬𝐢𝐨𝐧 KI verändert nicht nur, wie wir Aufgaben lösen – sondern vor allem, ob wir dabei lernen. 𝐃𝐚𝐬 𝐳𝐞𝐢𝐠𝐭 𝐬𝐢𝐜𝐡 𝐤𝐨𝐧𝐤𝐫𝐞𝐭 𝐢𝐧 𝐝𝐞𝐫 𝐈𝐧𝐭𝐞𝐫𝐚𝐤𝐭𝐢𝐨𝐧 𝐦𝐢𝐭 𝐊𝐈: - „Mach das für mich“ - also Aufgaben delegieren. - „Zeig mir die Logik dahinter – warum funktioniert mein Ansatz noch nicht?“ – um Lösungen und Konzepte zu verstehen. Im ersten Fall wird Arbeit abgegeben - mit Risiko zum Deskilling. Im zweiten erweitern wir unsere Kompetenzen. Wenn wir KI einsetzen, können wir über Tools und Inhalte sprechen. Wir müssen aber auch festlegen und messen, wie damit gearbeitet wird - wenn wir Fähigkeiten fördern wollen. 𝐒𝐨𝐧𝐬𝐭 𝐨𝐩𝐭𝐢𝐦𝐢𝐞𝐫𝐞𝐧 𝐰𝐢𝐫 𝐚𝐮𝐟 𝐆𝐞𝐬𝐜𝐡𝐰𝐢𝐧𝐝𝐢𝐠𝐤𝐞𝐢𝐭 – 𝐮𝐧𝐝 𝐯𝐞𝐫𝐥𝐢𝐞𝐫𝐞𝐧 𝐝𝐚𝐛𝐞𝐢 𝐅ä𝐡𝐢𝐠𝐤𝐞𝐢𝐭𝐞𝐧. Danke für die spannende gemeinsame Diskussion an Klemens Budde, Andrea Schmidt-Rumposch und Bjoern Eskofier sowie für die Vorbereitung an das PLS-Team Pia Schroeder, Christine Wirth und Thomas Steiner. #DMEA #KI #PatientCare #Gesundheitswesen #Medizin #Pflege #DigitalHealth #Ausbildung #Ottobock
-
George Pirvu gefällt dasMit Norman Hofer und seinem Bereich habe ich 4+ Jahre als Kunde zusammengearbeitet. Was uns immer verbunden hat: die Überzeugung, dass Data & AI nur dann Wirkung zeigen, wenn sie nachhaltig verankert sind – nicht als Buzzword, sondern als Arbeitsweise. Data & AI nachhaltig mit der Marke msg zu verknüpfen, daran arbeiten wir nun gemeinsam 🚀 Danke Norman, dass Du extra nach Wien gekommen bist, um mich willkommen zu heißen. Sagt einiges über die Kultur, in die ich einsteige. 🤝 msg Plaut Austria GmbH msg #data #ai #softwareGeorge Pirvu gefällt dasSeit 4+ Jahren arbeiten wir in der Kunde-Dienstleister-Konstellation zusammen. 𝗡𝘂𝗻 𝗮𝗹𝘀 𝗞𝗼𝗹𝗹𝗲𝗴𝗲𝗻! 🤝 Ich freue mich wahnsinnig, Dich bei msg begrüßen zu dürfen und gemeinsam mit Dir unsere #Data, #Analytics und #AI Präsenz in Österreich auszubauen. Ein Post, der so dermaßen von Herzen kommt lieber Daniel Wallner! 𝙒𝙚𝙡𝙘𝙤𝙢𝙚 𝙩𝙤 𝙢𝙨𝙜 🚀
-
George Pirvu hat darauf reagiertGeorge Pirvu hat darauf reagiertOh noes.. The Mythos Superior AI that Misanthropic Dario was too aftaid to release to public, got hacked.. So first Claude Code leaked its own code (blamed on 'human error' bs) and now it comes out Mythos is Greek for Swiss Cheese.. 😬 So much for AGI claims again. Or maybe they were prompting it all wrong.. The truth is its just another LLM trained with carefully curated, 'distilled', and labeled security code and tooling datasets.. Still drifts. Still only serves to increase your org's attack surface. Could be a cute PenTest tool. #ai
-
George Pirvu gefällt dasGeorge Pirvu gefällt dasEveryone talks about the future of industry. Here, it’s being built. Those who turn data into action will lead — the rest will follow. #HMI
-
George Pirvu gefällt dasGeorge Pirvu gefällt dasWhat makes Seed IQ so effective in Quantum Error Correction.. Most standard quantum error correction (QEC) is reactive by design. Surface code workflows, stabilizer loops, minimum weight perfect matching (MWPM) + union find (UF), all follow the same logic.. Measure syndromes, infer what happened, apply a correction. The system acts only after deviation or error has already happened. It reconstructs failure from what is left behind rather than governing evolution before instability compounds. That is the core limitation. If MWPM + UF were enough, fault tolerant quantum computing (FTQC) would already be here. They resolve manifested error but they do not maintain an ongoing forward hypothesis structure over evolving system state, and they do not determine whether the current evolution should be allowed to continue, be redirected, or left alone before decoherence amplifies. Seed IQ does. Its strength is not better post hoc decoding. Its strength is continuous sensing of the system, maintenance of concurrent hypotheses over possible forward evolution, and real time determination of whether the present trajectory remains admissible under the adaptive and fixed priors, and constraints governing coherence preservation. The priors matter because they define what valid evolution looks like, which transitions remain acceptable, which perturbations are tolerable, and which deviations indicate genuine departure from the goal. Not every deviation is something that should be touched. Some deviations or errors remain inside admissible evolution and will relax without intervention. Others are the early formation of decoherence and need redirection before they propagate. A standard reactive decoder does not really know that until the problem has already expressed itself as syndrome structure. Seed IQ is operating earlier than that. It is not waiting for full manifestation. It is maintaining hypotheses over evolving state and determining whether the system is still moving in a coherence preserving way or whether intervention has become necessary. So Seed IQ is not simply correcting errors. It is governing evolution. That is also where the multiagent advantage comes from, or the distributed coordination of Adaptive Multiagent Autonomous Control. Each agent is local in sensing and updating, but all agents operate under the same priors and constraints. They are not isolated local decoders, and they do not need heavy explicit message passing or consensus overhead. Coordination is implicit because they are all participating in the same admissible mathematical structure. Global coherence emerges from locally consistent belief propagation across the system. Standard QEC detects, reconstructs, and corrects. Seed IQ senses, maintains hypotheses, determines whether evolution remains admissible, and intervenes only when coherence preserving execution is at risk. Which is why it scales and performs so much more effectively. AIX Global Innovations #ai #quantum
-
George Pirvu gefällt dasFragmented data. Inconsistent reporting. Delayed decisions. Most organizations don’t have a reporting problem-they have a data foundation problem. At Adastra, we helped design and deliver numerous enterprise-scale data platforms that unify data across on-prem and cloud, enabling: A single, trusted source of truth Faster, decision-ready insights Scalable self-service analytics Reduced reliance on legacy systems The impact is tangible-better decisions, faster. Because real transformation doesn’t start with dashboards. It starts with getting the data right. #DigitalTransformations AdastraGeorge Pirvu gefällt dasAt Volkswagen Group IT data was distributed across multiple systems – both on-prem and in the cloud. Reporting required manual consolidation, which slowed down access to insights and made it harder to work with consistent data. Adastra implemented a centralized IT4IT platform that connects data ingestion, processing, and reporting into one framework. Under the hood: - Azure-based architecture combining cloud and on-prem data - Azure Data Factory pipelines orchestrating data ingestion and transformation - Power BI Service for reporting and self-service access - Governance with role-based access and audit capabilities What this enabled: ✔️ consistent data across reporting ✔️ faster report delivery ✔️ access to data for non-technical users ✔️ reduced dependency on on-prem infrastructure A good reminder that reporting is not just about visualization – it’s about how data is integrated and prepared upstream. 👉 Full story: https://lnkd.in/dQbZkfsw
Berufserfahrung und Ausbildung
-
Adastra
Gesamte Berufserfahrung von George Pirvu anzeigen
Jobbezeichnung, Beschäftigungsdauer und mehr ansehen.
Wenn Sie auf „Weiter“ klicken, um Mitglied zu werden oder sich einzuloggen, stimmen Sie der Nutzervereinbarung, der Datenschutzrichtlinie und der Cookie-Richtlinie von LinkedIn zu.
Bescheinigungen und Zertifikate
-
-
-
-
-
-
Certificate Implementing a Data Warehouse with SQL Server
Microsoft Virtual Academy
Ausgestellt: -
Reimagine Finance BI for Sales Organizations
Microsoft Virtual Academy
Ausgestellt: -
Java Fundamentals
Cisco Academy
Ausgestellt: -
-
Kurse
-
DACHSER Corporate Compliance
-
-
Dachser Academy Controlling Course
-
-
German Language
B2
Sprachen
-
English
Fließend
-
German
Fließend
-
Romanian
Muttersprache oder zweisprachig
-
Spanish
Grundkenntnisse
Organisationen
-
StackOverflow
-
–Heute
Erhaltene Empfehlungen
1 Person hat George Pirvu empfohlen
Jetzt anmelden und ansehen