About
I take AI from impressive demo to systems the business can actually measure. That…
Services
Articles by George
-
How to Keep Your Data Safe with Claude Cowork: The Risk Your License Won't Fix
How to Keep Your Data Safe with Claude Cowork: The Risk Your License Won't Fix
1. Introduction Claude Cowork is quickly becoming one of the fastest-growing AI tools for enterprise work.
-
How to manage Machine Learning and Data Science projects
How to manage Machine Learning and Data Science projects
Machine learning (ML) and Data Science (DS) projects are hard to manage. Because projects are research-like in nature…
15 Comments
Activity
7K followers
-
George Bezerra, PhD shared thisMy latest article was just published in CIO: “Not Every Problem Needs an AI Agent.” After spending years building machine learning systems and, more recently, working extensively with LLMs and agentic systems, I’ve become increasingly convinced that one of the most important AI engineering skills is knowing when not to use AI. In the article, I look at four examples where the right answer was different: → A recommender system where GenAI simply wasn’t the right tool → Content moderation where ML and LLMs each had a distinct role → Database querying where deterministic logic beats probabilistic SQL generation → Search where an LLM solved a cold-start problem but didn’t belong in production The common thread is simple: Use the simplest technology that meets the requirements. Sometimes that’s deterministic software. Sometimes it’s ML. Sometimes it’s an LLM. And sometimes it really is an agent. The goal shouldn’t be to maximize the amount of AI in a system. It should be to maximize the value the system delivers. Read the full article in CIO: https://lnkd.in/gdejm_aX #AI #ArtificialIntelligence #MachineLearning #AIEngineering #AgenticAI
-
George Bezerra, PhD posted thisI have friends who are still dazzled by what AI can do. It takes time to realize that there is no miracle. To make AI work in the real world it requires a lot more scaffolding and hand holding. This is why it is often said that dazzling AI demos often fail in production. It is easy to build something impressive, but building something that adds value at scale is where the work is.
-
George Bezerra, PhD shared thisI'm excited to be speaking about AI agents in production at Outer Join, Metabase's first conference, this October here in Denver. If you're in the area, I'd love to see you there!George Bezerra, PhD shared thisYour AI agent demoed beautifully, but then it hit production. 😿 At OUTER JOIN, George Bezerra, PhD, will explain why the gap between demo and production usually isn't a model problem, but a data problem. Investigate an agent fielding executive questions against live business data, and learn what made answers trustworthy. Come hear George and develop your own AI data skills at OUTER JOIN in Denver, CO, from October 5-7! And if you have an interesting Metabase use case, an AI success story, or insights on running Metabase at scale, consider submitting a talk proposal by August 10. 🎤 https://lnkd.in/dJtVKgA2
-
George Bezerra, PhD posted thisConversational analytics is being sold by every major data platform. Here's why it can't deliver on its promise. Databricks has Genie. Snowflake has Cortex Analyst and Snowflake Intelligence. Microsoft has Copilot in Fabric and Power BI. The pitch is identical across all three: your business users ask a question in natural language and get an answer, no analyst in the middle. Notice who is actually using them. Not executives. Not business users. Mostly analysts, the exact people the product was supposed to make unnecessary, checking the SQL before anyone acts on it. Self-service requires a level of trust these systems cannot deliver. On enterprise-grade benchmarks, the leading agent framework solves 21.3% of tasks, against 91.2% on the academic ones. [Spider 2.0 (Lei et al., ICLR 2025)] But accuracy is not really the issue. The issue is that a business user cannot tell which answers are wrong. An analyst catches a bad join. A VP makes the wrong decision. What is missing was never a tool. It was expertise. Data models, semantic views, an MCP server, an agent: these are components of a system, and a system only works when someone who has built one before decides how the pieces fit and what human process runs alongside them. That judgment is the scarce thing. It always was. AI raises the stakes on getting it right, because now the wrong answer arrives fluently and instantly, with no one in the loop to catch it. You can buy every component from a single vendor tomorrow. Nobody sells you a system you can trust. If you have one of these running unsupervised in front of executives, I would like to hear about it. I do not think many people do.
-
George Bezerra, PhD shared thisI frequently get the same question from business leaders adopting Claude Cowork: "How do I keep my data safe?" Understanding your Anthropic license is a good start, but it is only part of the story. Some of the biggest risks, like cross-contamination (where one client's data ends up influencing another client's work), aren't solved by your subscription alone. Here are two things worth knowing: • The chat's Feedback Button can send conversation data back to Anthropic to help train models, even if you're on the Enterprise license. • Claude Cowork is not HIPAA compliant under any standard license. Even if you have a HIPAA-enabled Enterprise agreement, Cowork itself is not covered. I wrote a practical guide that breaks down how Claude Cowork handles your data, what each license tier actually protects, and where the remaining risks are.How to Keep Your Data Safe with Claude Cowork: The Risk Your License Won't FixHow to Keep Your Data Safe with Claude Cowork: The Risk Your License Won't Fix
-
George Bezerra, PhD shared thisIf you have a consumer Claude account and you're worried about what happens to your personal data, here's one setting you should change immediately: turn off model training. On consumer Claude accounts (Free, Pro, and Max), the setting to help improve Anthropic's models is enabled by default. That means your conversations may be used to train future versions of Claude. Why does that matter? Large language models learn statistical patterns from the data they're trained on. They don't store conversations like a database does. But because modern models contain billions of parameters, they can occasionally memorize parts of their training data. While it's unlikely that Claude would reproduce your conversations, there is no way to rule that possibility out completely. The good news is that avoiding this is simple. Go to Settings → Privacy → "Help improve our AI models" and turn it off. If you're using Claude for personal writing, financial information, legal documents, or anything else you'd rather keep out of future model training, there's little downside to disabling it. One final point: this setting only applies to consumer accounts. Enterprise plans have different data handling policies and, by default, do not use customer data for model training. Privacy isn't just about what you share. It's also about knowing which defaults you never meant to accept.
-
George Bezerra, PhD shared thisMost people worry about putting sensitive data into AI. The bigger risk is often the data that doesn't look sensitive at all. Companies are using Claude, ChatGPT, and Gemini every day to draft documents, analyze data, summarize meetings, and automate increasingly complex workflows. But many still aren't sure where the line is: what's safe to share, and what isn't. Some organizations push ahead, hoping their enterprise license protects them. Others limit adoption, worried about exposing confidential information. One fears moving too fast. The other fears falling behind. The reality is that most data incidents don't happen because the AI failed. They happen because someone gave it data they didn't realize carried legal, contractual, or regulatory obligations. Before asking whether an AI tool is safe, ask a simpler question: What kind of data am I giving it? Every piece of data carries an obligation. It might be none at all. It might be company confidentiality, a client NDA, a privacy regulation like GDPR or HIPAA, or a vendor license agreement. Those obligations determine the rules you have to follow, regardless of which AI platform you're using. The framework below is a simple mental model I use to classify enterprise data before it ever reaches an AI system. The risks that cause the biggest problems are rarely the obvious ones. They're the ones you didn't realize were there.
-
George Bezerra, PhD shared thisAI-first is the wrong strategy for most companies. Not because AI doesn't work. Because the math doesn't. There is a cost to autonomy that almost never makes it into the business case. The license and the implementation are the costs everyone counts. They are rarely the biggest ones. Here is what usually gets left out. → Maintenance. An automated workflow is not finished when it ships. Models change. Systems change. Business processes change. Someone has to keep it working. This cost never ends, and it grows with every workflow you automate. → Risk. When you remove the person, you remove the one who catches the mistakes. Now you are paying for the errors that reach your customer, or paying engineers to prevent them, or both. → Oversight. In practice, most companies do not eliminate the human. They turn them into a reviewer. That work is real, and it rarely shows up in the plan. → Headcount. You replace coordinators and analysts with platform engineers and AI engineers. Fewer people, higher salaries, harder to hire. → Opportunity cost. The engineers building and maintaining the agent are not building your product. Whether all of that is worth it depends on your situation. Volume matters, because these are mostly fixed costs, and a workflow that runs ten times a year will never pay them back. Variability matters, because work that is different every time resists automation. Rate of change matters, because a process that changes every quarter means maintenance forever. Industry and regulation matter, because they determine what an error costs. Which is why ROI does not always increase with autonomy. For some workflows, it does. For many, it peaks somewhere in the middle. And for some, the highest-return solution is a trained person using AI with little or no automation at all. The only way to know where you land is to do the math. Most companies never do. They estimate the benefits once, approve the business case, and never measure the ongoing costs that determine whether the investment actually pays off. Do the math first. It is cheaper than the wrong strategy.
-
George Bezerra, PhD shared thisNot every problem needs an AI agent. Yet under pressure to "adopt AI," I keep seeing teams reach for LLMs and agents when deterministic systems or predictive ML would solve the problem better. If you've built ML systems before the generative AI wave, the pattern is familiar. You pay for it three ways: → Higher cost → Higher latency → Less predictable behavior ...for capabilities the problem never required. The answer isn't less AI. It's choosing the simplest system that solves the problem. Think of AI as a continuum: Deterministic systems — The behavior is defined by code or explicit logic. Rules, workflows, SQL, APIs, and traditional software. Fast, inexpensive, and highly predictable. Predictive ML — The goal is forecasting, scoring, ranking, or classification, and you have labeled data. Statistical learning solves these problems efficiently. LLMs — The task requires language understanding, reasoning over unstructured information, or content generation. Agents — The task requires planning across multiple steps, tool use, and adapting based on intermediate results. Every step to the right buys you more flexibility. Every step also increases cost, latency, and uncertainty. So don't ask: "Where can we use an agent?" Ask: "What's the simplest system that solves this problem?" The best AI systems aren't the ones using the newest technology. They're the ones using the simplest architecture that meets the requirements.
George Bezerra, PhD commented on a post
1w
Enjoy this new phase, David. Wishing you all the best!
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisAmazon is shutting down Mechanical Turk next month. It makes sense. Horizontal marketplaces for unskilled human tasks are being squeezed out by AI. I wonder if we'll ever see another human task marketplace like it. Perhaps for specialized training-data work (i.e., domain expertise rather than unskilled tasks). https://www.mturk.com/
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisHappy to share that Timothy Chi has joined the Giftory Board 🥂 Tim co-founded WeddingWire, scaled it into The Knot Worldwide, led it as CEO for nearly two decades, and continues to serve on the Board of Directors. He built the category-defining marketplace for the moments people plan their whole lives around while we're building the category-defining brand for every moment in between: birthdays, anniversaries, Christmas, thank-yous, and just-because. Giftory offers over 5,000 curated experiences. Cooking classes, supercar drives, dining cruises. Gifts you can send as a personalized digital voucher or a gift box that shows up at the door. Having Tim in the room will be a game changer as we aim to hit $100m in sales next year while delivering a best in class customer experience. And since we are neighbors in DC, I am personally thrilled to benefit from his expertise in consumer internet & online marketplaces, as well as his CEO mentorship. Welcome aboard Tim!
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisLast night was so much fun - I got to share Local Blooms at the Southern Colorado Startup Showcase. We played "Guess which bouquet was grown here?", gave out mini-bouquets, and talked about how important it is to change the way we buy flowers in the U.S. Even Mayor Blessing ‘Yemi’ Mobolade played along! I'm really grateful to live in a community that has so much support for start-ups and small businesses and especially grateful that our mayor cares so deeply about supporting local small businesses - it means so much. Thank you to Exponential Impact, El Pomar Institute for Innovation and Commercialization (EPIIC), Catalyst Campus, Colorado Springs Emergent Campus, and City of Colorado Springs for putting on this event and bringing together so many bright new companies from our community - it was truly a special evening and I'm so grateful I got to be a part of it.
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisWe weathered a storm! We saw a rainbow and a full moon! We kept it 💯 at the latest edition of the Denver Data Dudes and Dudettes Meetup! Kent Lindquist Neil Fonseca, CFA Kamal 🚀 Maheshwari Kimberly Southerland Leah Kabira Jonathan "Iggi" Sharley Ian Ellis, PhD Alex Lopez Hartley
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisHow do you compare countries without pretending a single ranking can tell the whole story? I went out to try to answer that question and today I’m making Global Systems Assessment publicly available. It started as a way to compare how countries perform across the systems that shape everyday life, but it has grown into a broader research project. The project looks at areas like institutions, human capital, economic capacity, stability, and system complexity, with the goal of making the reasoning and evidence behind each assessment easy to inspect. I did not want this to be a black-box ranking system. The site includes the country atlas, methodology, research notes, evidence reviews, failed tests, data limitations, and the underlying framework used to generate the scores. The methodology, data pipeline, documentation, and website are all public. It is still evolving, and that is part of why I wanted to publish it now. I would like people to test the assumptions, point out gaps, challenge weak conclusions, and help make the framework better over time. Explore the website: https://lnkd.in/gTFUMugP View the repository: https://lnkd.in/g3yAquZE If you work in public policy, international development, data journalism, governance research, political science, or anything adjacent to those fields, I’d genuinely value your perspective. #PublicPolicy #DataScience #Governance #InternationalDevelopment #OpenResearch
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisSoCo Start up week kickoff with the best! Cari Hanrahan, Rachel Kovacs. Thanks Jon-Michael Horst, Rachael Maxwell, Max Cupp, Coral Laski, yall are awesome!! Exponential Impact, Southern Colorado Startup,
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisWho wants to help measure and optimize Rivian's next evolution of marketing with me?? All in the spirit to keep the world adventurous forever.
-
George Bezerra, PhD liked thisGeorge Bezerra, PhD liked thisAfter almost five years, my chapter at The Knot Worldwide has come to an end. It has been quite a journey. I’ve had the opportunity to work with incredibly talented people, lead great teams, solve some very difficult problems, and learn a lot along the way — about technology, leadership, business, and myself. I leave proud of what we built together and, above all, grateful for the people I met and learned from during these years. Thank you to everyone who was part of the journey. 💚 As for what’s next… for once, I don’t have an immediate answer — and I’m actually excited about that. Those who know me well know that I’ve always been an entrepreneur at heart. So I want to give myself a few months to reconnect with that side of me: explore ideas, learn, build things, and see where curiosity takes me. And it feels like a particularly exciting time to do it. Technology is going through one of the biggest transformations I’ve seen in my career, and AI is making things possible for small teams — and even individuals — that would have been incredibly difficult just a few years ago. Maybe this exploration leads me to build something of my own. Maybe it takes me somewhere completely different. For now, I’m deliberately leaving that question open. And if you’re working on an interesting problem where AI could fundamentally change how things are done — or you have an idea you think is worth exploring — I’d love to hear from you. No promises, no pitch deck required. Just a coffee and a good conversation to see where it takes us. ☕️🚀 On to the next adventure!
Experience & Education
Honors & Awards
-
Excellence in Graduate Research
Sigma Xi, The Scientific Research Society
Languages
-
English
-
-
Portuguese
-
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content