Ebrahim Elgaml

Journey to 1000 models from Meta. It explains how they manage so many models in production simultaneously. The key idea is observability and automation. System Design 101. Worth reading.

118

1 Kommentar

Heading to RecSys '25 next week with the Zalando team. I'm excited to share two contributions from our team that will be presented at the conference: ✨ Efficient and Effective Query Context-Aware Learning-to-Rank Model for Sequential Recommendation https://lnkd.in/e3cYbsDc This paper dives into our research on integrating query (target) context into a transformer model. We explore new approaches and quantify their effectiveness under industrial feature store constraints. ✨ Short-Form Video Recommendations with Multimodal Embeddings: Addressing Cold-Start and Bias Challenges https://lnkd.in/e8Pw-27Q This short format paper highlights why conventional recommendation methods often fall short for short-form video. It explores the complexities of integrating video into an e-commerce ecosystem and showcases the use of multimodal embeddings. These papers were a collaborative effort with amazing colleagues (in alphabetical order): Alisa Mironenko, Egor Malykh, Evgeny Labzin, Francesca Catino, Katya Mirylenka, Maarten Versteegh, Marco-Andrea Buchmann, Marjan Celikik, and Vladimir Vlasov. Looking forward to a week of great discussions and connecting with the recommender systems community!

42

2 Kommentare

Want to learn data science, machine learning, and AI and happen to live in or around Zurich? Apply to ETH Zürich's DAS (Diploma in Advanced Studies) Data Science! I recently completed the DAS program after two years of part-time study. The DAS allows you to choose from a range of master's-level classes across data science and adjacent disciplines (computer science, mathematics, statistics, neuroscience, and electrical engineering). There are over forty classes to choose from. I managed to squeeze in eight classes over the two years, all of which were uniformly excellent. The teaching was among the best I have experienced, and the courses were incredibly up-to-date, with us often studying cutting edge material that was only released in the preceding weeks and months. The assignments were also well thought-out and instructive. Going into this program, I already had a solid foundation in machine learning, but the diploma allowed me to fill in many gaps and to spend much more time and effort investing in areas I was less familiar with: reinforcement learning, Bayesian machine learning, and more esoteric neuroscience-based approaches. Even in the areas that I was already familiar with, there was plenty of learning to be had from large language model watermarking, fundamentals of natural language processing theory, and a more statistically rigorous view of model ensembling, to give a few examples. Another awesome benefit of the program is the networking. There are other part-time professional students like myself, on top of the usual ETH master's and PhD students, so I got to network with, collaborate with, and become friends with many awesome professionals and students. My only criticism of the program is that I would happily do even more. I would eagerly do another few years and earn another set of ECTS credits to get an MAS (Master of Advanced Studies) if it were offered. Hint, hint, ETH Zurich | School for Continuing Education ;) I have done a lot of part-time study over the years and I can state that this DAS was my favorite so far. Highly recommended for those with any interest in the field! Diploma information: https://lnkd.in/eDKwtMCZ Class list: https://lnkd.in/ehutbZ-g

210

13 Kommentare

At this year’s PyConDE & PyData 2025 in Darmstadt, I had the chance to talk about "From Algorithm to Action: Building a DIY Distributed Trading Platform with Open Source". In the talk, I walked through key challenges, design decisions, and lessons learned from real-world trading systems. I shared practical tips that can help you to build a modern trading system yourself The recording is now available on YouTube:🎥 https://lnkd.in/dwjWQNJC #Software #PyCon2025 #Freelance #SoftwareArchitecture #Trading

20

What does an average mean to you? Averages are meant to represent a distribution, but what happens when your use case is a minority within that distribution? Let’s take a concrete example: LLM benchmark results are often reported on datasets that are predominantly English or Chinese, where languages like Arabic are underrepresented. (Approximately 3–4% of web content is in Arabic, despite its global significance.) As someone developing for the Arabic-speaking community, you should take these benchmarks with a grain of salt. This doesn’t mean dismissing them entirely, but it’s critical to consider their limitations. So, how should you choose your models moving forward? Start with top-performing models. Evaluate them against your own benchmark, tailored to your specific needs. This would be far easier if we had a reliable, up-to-date Arabic benchmark—but many existing ones are incomplete or outdated. Until then, choose your models your way. Test rigorously, prioritize your own use-case, and don’t let generalized averages dictate your decisions.

4

🚀 Just wrapped up a deep dive into Pydantic for LLM workflows – When building LLM-powered systems, one of the biggest challenges is ensuring data is clean, consistent, and validated before feeding it into workflows. That’s exactly where Pydantic shines. Why Pydantic is critical for LLM projects: 🔹 Data validation at scale – Automatically checks input/output structures, so you don’t have to manually catch errors. 🔹 Schema definition – Clearly define the shape of your prompts, responses, and metadata using Pydantic models. But one of the most powerful capabilities is how Pydantic acts as a feedback loop for the model itself. 👉 Imagine this: You expect the LLM to return a JSON with fields like title, summary, tags. But the model returns something incomplete, misformatted, or with the wrong types. Normally, you’d get silent failures or messy downstream errors. With Pydantic models, the response is immediately validated against a strict schema: ❌ If it fails → Pydantic throws a structured error. 🔁 That error can be sent back to the LLM as feedback, prompting it to regenerate the response correctly. This creates a self-correcting loop, where the LLM keeps refining until the output meets your schema. Why this matters in LLM projects: 🔹 Consistent outputs → Models stay aligned with your workflow requirements. 🔹 Reduced debugging → Errors are clear and structured, not hidden in messy text. 🔹 Scalable workflows → Works across multi-agent systems and production pipelines. 💡 In essence: Pydantic transforms “LLM guesswork” into predictable, validated results – and gives models the chance to fix themselves. Have you tried using Pydantic as a feedback mechanism in your LLM workflows? Would love to hear your thoughts 👇 #AI #LLM #Python #Pydantic #LangChain #AIEngineering

15

3 Kommentare

Very interesting insights yesterday at the MLOps Munich Meetup hosted by Slalom. (1) Long running low-cost inference pitched by Doubleword's Meryem Arik were an interesting and new approach (at least for me) at the general problem of reducing token cost (2) Current AWS features regarding RAG and knowledge retrieval with less friction and manual implementation overhead than the common approach were a good impulse by Pradeep Gunana as well. Looking forward to the next Meetup ✌

10

1 Kommentar

Interest in public markets in the region is on a rise, more and more companies are getting listed everyday. Visibility on these still remains unsolved. Just launched Premargin with Zubair Naeem Paracha, a simple way to explore public company data across Saudi, UAE, Qatar, and Kuwait. Still early, but already useful. Would love feedback if you try it. https://www.premargin.com/

26

3 Kommentare

Hi there! It took me a while to find my true passion: Software and ML engineering. That’s why I’m hyped to start as an ML Engineer at ML6 | Your partner in AI, a company at the forefront of AI innovation. A big thank you to Matthias and Laurence for the warm welcome 🧡 Let the programming begin! PS: I’m interested in AI for Good 🌍. If you have ideas on how we can use AI to make the world a better place, I’d love to connect and chat! #MachineLearning #ML6 #AIForGood

202

48 Kommentare

Building AI products at scale gets expensive fast, and we've learned that firsthand building Sift Skills. One thing that's made a real difference for us is prompt caching. If you're building products that require frequent, but unchanged, LLM API calls, it's worth understanding how it works. Check out the blog by Nick DeJesus below👇🏾

12

Can't make it to #ACL2025 this year, but for people interested in RL for factuality and textual decomposition, please check out our paper: https://lnkd.in/ed3zPXfY TL;DR: We found a mismatch between the decomposition policy and LLM verifier, and propose a dynamic training paradigm to bridge the gap. This is our teaser video 😀 https://lnkd.in/epQukV8p

39

3 years ago, I got the chance to speak at Google #Berlin MLOps meetup. It was one of those moments where you speak, but you also deeply listen. And here’s what I walked away with 👇 MLOps is misunderstood. ✖️ AWS SageMaker ✖️ Databricks ✖️ MLflow ✖️ Kubernetes ✖️ A solo engineer’s job ✖️ Just for “MLOps Engineers” ✖️ A one-size-fits-all solution ✅ MLOps is a mindset. A team sport. A system of principles, practices, and cultural shifts that make ML actually work in production. And even now in the era of LLMs, RAG, and Agents —> 👉 MLOps still remains deeply relevant in AI Engineering. Because without: 🔁 Reproducibility 📦 Automation 📊 Monitoring 📈 Scalability 🧱 Reliability …your shiny GenAI demo won’t survive a week in prod. If you’re serious about building real-world AI systems, you need to care about more than just the model. 📌 𝗬𝗼𝘂 𝗻𝗲𝗲𝗱 𝗠𝗟𝗢𝗽𝘀. Do you agree? ♻️ Repost if this resonates with you 💙 -- We’re launching “𝗧𝗵𝗲 𝗠𝗼𝘁𝗵𝗲𝗿 𝗼𝗳 𝗔𝗜 𝗣𝗿𝗼𝗷𝗲𝗰𝘁” - Starts with RAG, expands to Agents, RecSys, MLOps, and more. Interested? Register here - https://lnkd.in/dBVBdkEG

181

19 Kommentare

ANOTHER HUGE SIGNAL OF WHERE THE TECHNOLOGY MARKET IS HEADING Yann LeCun, former Chief AI Scientist at Meta and one of the pioneers of modern deep learning, has just raised $1.03bn for his new startup, Advanced Machine Intelligence (AMI) in what is being reported as the largest seed round ever raised by a European company. The Paris-based company is already valued at around $3.5bn pre-money, with backing from some serious names including Nvidia, Bezos Expeditions, Cathay Innovation and Temasek. What’s interesting is the direction they’re taking! Rather than simply scaling large language models, AMI is focused on building “world models” AI systems designed to understand how the physical world behaves, learning from video and spatial data so they can reason, plan and interact with real environments. Think robotics, manufacturing, autonomous systems and real-world automation, not just chatbots. For those of us working across software engineering, cloud infrastructure and testing, this kind of investment is another reminder of how quickly the market is evolving. The companies that win in the next decade won’t just build software. They’ll build systems that understand and interact with the real world. And that’s going to require some very interesting engineering talent. Exciting times ahead!!! #Tech #TechRecruitment #PRS #ArtificialIntelligence #AI #SoftwareEngineering #CloudComputing #MachineLearning #TechStartups

36

3 Kommentare

OpenAI released open-weight models last week. They run locally. They’re cheap. But they don’t outperform Qwen, GLM, or even Deepseek. So why did this release matter so much? And what does it change if you’re building with LLMs? #ai #openai

22

2 Kommentare

I’m excited to share a very special milestone: our paper “Comparing Static Analyses for Improved Semantic Conflict Detection” has been accepted for publication in Automated Software Engineering. At a time when Artificial Intelligence is reshaping the entire software development landscape, the challenges of keeping code consistent, reliable, and free of subtle errors are growing as well. Even with increasingly intelligent tools, semantic conflicts remain one of the most complex and least supported problems during merge processes, often unnoticed by modern version control systems. This paper shows that by combining different static analysis techniques, we can make significant progress in detecting such conflicts and uncover scenarios that neither generative AI nor traditional tools can currently identify. #staticanalysis #soot #semanticConflicts #Git #AI #algorithms #scala #elixir #mergeScenarios #research

36

1 Kommentar

Berlin data folks, this one’s for the builders. We’re joining Delivery Hero and Zalando to talk real-world Apache Flink in production — not theory, not hype, but the work that actually moves metrics, cost, and platform performance. From large-scale metric computation to cutting state by 75%, to reducing data platform costs through continuous materialization — this is what production streaming looks like when it’s done right. If you’re working with streaming systems, optimizing Flink, or deciding between Table API and DataStream, you’ll want to be in the room. See you in Berlin. 🍕⚡ Where production systems get discussed by the people who run them. #ApacheFlink #DataEngineering #StreamProcessing #Berlin

6

We once had a ML model with great AUC (a type of model accuracy) Everyone was happy. Until the business team asked: “Can we trust the 0.85 risk score for this user?” The natural response comes - of course, this model has a great AUC. Well, that’s just one part of the story. Here’s how: AUC tells you how often a risky user gets a higher score than a non-risky one. It measures ranking quality. But it says nothing about whether a predicted 0.85 actually means 85% chance of fraud. It could very well happen that predictions between 0.8–0.9 are only 60% accurate. Still confused? - Say, you took all the users having predictions b/w 0.8-0.9, and checked their actual fraudulent flag. Saw only 60% of those high predicted risky users were fraudulent. Here, your model is over-confident. - Still in denial!! You now picked all the users having predictions b/w 0.3-0.4 and these users had fraudulent flag only 50% of the time. The model is under-confident there. Your whole world gets shaken, but hey wait, that’s where “calibration” comes for your help. Calibration reshapes predicted probabilities so that they reflect reality more closely. Before calibration: - Model predicts 0.90 → but only 60% of such users are actually fraud - Model predicts 0.40 → but 50% turn out to be fraud So the model is: - Over-confident in high-risk cases - Under-confident in medium-risk cases After calibration, those probabilities might adjust to: - 0.90 → 0.68 - 0.40 → 0.52 Notice something important: The ranking between users hasn’t changed. The highest-risk user is still ranked highest. Use it when your system: • Uses probabilities for decisions (risk-based approvals, churn interventions) • Allocates budgets based on predicted risk • Feeds probabilities into downstream cost models • Automates actions using thresholds In short — when the number itself matters. When might you not need it? If: • You only care about ranking (e.g., top-K recommendations) • You’re using the model strictly for ordering, not probability interpretation • Thresholds are continuously optimised downstream #data #MachineLearning #datascience

52