Mehdi Mujtaba

Quantization in Large Language Models (LLMs) As LLMs scale, intelligence is no longer the bottleneck — memory, latency, and cost are. Quantization turns out to be one of the most powerful (and misunderstood) techniques enabling real-world deployment of large models. In this blog, I break down: • Why LLMs are surprisingly quantization-friendly • INT8 vs INT4 vs NF4 — what actually works in practice • How modern architectures absorb low-bit noise • Where quantization fails and how to design around it Quantization is no longer an optimization step — it’s foundational infrastructure for production LLM systems. Link: https://lnkd.in/gqGje75G #LLMs #AIEngineering #Quantization #MachineLearning #GenAI #MLOps

9

2 Comments

We're hiring three AI Analytics Engineers on the Data Science & Analytics team at Airtable. These are early to mid-career roles for builders who are excited about working at the intersection of data science, analytics, and AI. One of these roles will report directly to me as we spin up a new AI & Analytics Platform function. Things are moving fast. The most impactful work in data right now isn't building models — it's fundamentally changing how a company gets answers from its data. That's where my focus is, and it's why Airtable is investing in a new team to own it. Over the past six months, our team stopped experimenting with AI on the side and restructured around it: - Agent-powered development workflows — I built shared AI systems for the data team embedded in how we write SQL, build pipelines, document business logic, and debug. Team members are regularly seeing 2x to 10x speed improvements — not from shortcuts, but from AI woven into every step of how we build. - Conversational analytics — Ad-hoc questions that used to fill our queue from leadership, product, GTM, and finance are increasingly answered by an AI agent — in natural language, where people already work. -A context layer that makes it trustworthy — We built a system that gives the agent real business knowledge — which table is correct for a given analysis, how metrics differ across teams, why something spiked last quarter. Context that usually lives in an analyst's head. - A learning loop that compounds — Every user correction becomes a permanent improvement, not a one-off fix. — We're doubling down. Airtable is standing up an AI & Analytics Platform function within DS&A to keep building at the intersection of data science, ML, analytics, and AI — and I'm leading it. We need more builders. AI Analytics Engineer, AI & Analytics Platform First hire on this team. You'll build the context layer that makes AI tools accurate, design evaluation frameworks that make them trustworthy, and develop agent systems that let the org self-serve with confidence. You'll work directly with me to shape what this function becomes. https://lnkd.in/gFstQqQb We're also hiring two roles embedded in other parts of the business: AI Analytics Engineer, Marketing Analytics Own data infrastructure, dashboards, and AI-native tooling for Marketing. Build the canonical data sources leadership relies on. https://lnkd.in/gYHS55ma AI Analytics Engineer, Business Analytics Own financial data models powering ACV, ARR, billings, and revenue. Partner with a Finance team that's building their own AI agents and workflows. You'll build the data foundation they run on. https://lnkd.in/gV5hnJjG — I'm looking for people who build before they're asked. Who have a side project that started as "I wonder if I could..." and turned into something real. Who pick up a new tool and immediately try to break it. If that sounds like you — reach out or apply. I'd love to talk.

79

8 Comments

I Made It Up new feature drop: Analysis. This is a new tool that processes synthetic interview transcripts for quantitive (numbers) and qualitative (stories) analysis: • Quantitative - means + variance + statistical comparisons across demographic groups + outlier detection • Qualitative - theme extraction + clustering into higher-level patterns + supporting quotes Insights before you're back with coffee. If you don't know: I Made It Up is a synthetic research toolkit that plugs straight into AI chat like ChatGPT and Claude. (If that's jargon, 'synthetic research' is methodology that uses AI to simulate digital personas as avatars to explore how people might think. Made up? Yes. Useful? We think so.) Big words. What's it like in practice? Say you interview 36 personas about a product idea. Your AI then passes the transcripts to the Analysis tool for processing. You + AI then get results back that enable further analysis + thinking. Quantitative tells you: "Gen Z rated this 2.1 out of 5, while Millennials gave it 3.8 - that's statistically significant with a large effect size". Qualitative tells you: "Three themes emerged - price sensitivity, trust concerns, and feature confusion - here are the quotes that back each one up". Super early stage. Interviews are limited to six questions for six personas. This goes up soon and when it does, the Analysis tool becomes indispensable. Want to know more? If you're already using it, just ask ChatGPT or Claude what you can do with I Made It Up's Analysis and Retrieve tools. If you're not yet, check Getting Started on website linked in comments. Available on Pro accounts and above. Lots under-the-hood to unpack. Blog soon. Questions? Hit me up. Would love to talk and if you're local, can drop in to your office for a chat. 

7

2 Comments

I often hear from early-career ML/DS folks who feel stuck or from folks who want to move in AI/ML. Here’s what I’d recommend based on my own hiring experience: 💻 Reframe your resume - One page only: Focus on the last 1–2 years of most relevant work. If >10 years, add second page. Make sure you can explain each and every project/ algo in detail. - Bullet points over paragraphs: “Built and deployed an anomaly detector in PyTorch on 1M records with 95% precision” beats “Experience with PyTorch.” - Group courses vs. projects: Put certificates (DeepLearning.AI, MLOps, GCP) in a small “Training” section; highlight actual projects and results first. This can also go into the skill section. Make sure whatever you put, you've actually worked on. 💻 Show production readiness - Recruiters want more than “I trained a model.” - Emphasize (learn if needed) end-to-end work: data pipelines, Docker containers, CI/CD, cloud deployments, monitoring. 💻 Build a portfolio of small wins - Pick 2–3 mini-projects that mimic real problems: train a neural network, RAG chatbot, data ETL pipeline, simple API serving a model. Perhaps share your experience on medium/blog. - Share code on GitHub with a clear README and hosted demo (e.g., Streamlit). Create PR requests to open source projects, if possible. 💻 Ace coding challenges - Practice timed Python/data-structure problems on LeetCode or CodeSignal. Can't emphasize how important and useful this is, purely for interview. - Practice again. Even when not interviewing. 💻 Network with purpose - Ask for 15-minute “resume reviews” or quick feedback on a project. - Engage on Discord communities (e.g., MLOps, DataTalks) — personal referrals carry weight. Landing that first ML role can feel like an uphill climb, but every resume tweak, project post, or friendly referral moves you closer. Keep iterating—and you’ll break through.

308

24 Comments

Agents and “agent claws” are often framed as AI taking over entire sectors. I think that’s the wrong mental model. A better model is mentored autonomy! Think of a strong junior engineer or analyst: You give them clear scope the right context access to a bounded set of tools and you evaluate their work continuously, keeping what’s good, reverting what’s not That’s exactly the pattern I take from https://lnkd.in/eZxedMDa: you’re not “outsourcing research.” You’re defining a small but real loop where an agent can iterate autonomously, inside constraints, and you wake up to a log of attempts and improvements. Autonomy is valuable only when it is bounded. Context is an engineered input, not a magic window. Capabilities are permissions, not features and evaluation is not a gate, it’s the supervisor! Use agents the way you leverage specialized resources: delegate deliberately, constrain responsibly, and review relentlessly! #AgenticAI #AgenticEngineering #AutoResearch #AIEngineering #HumanInTheLoop

20

3 Comments

How a 7B Model(SLMs) Beat GPT-4o(LLMs) in Agentic Workflows 🤯📉 We often assume that bigger is better when it comes to LLMs, especially for complex reasoning tasks. In my latest video, I break down AgentFlow, a new framework where a 7B parameter model outperforms top-tier proprietary models (like GPT-4o) on complex, multi-turn agentic benchmarks. Here’s what I cover in the deep dive: 🔹 From Monolithic to Modular: Most tool-using LLMs try to do everything (reasoning + tool calling) in one monolithic process. AgentFlow decomposes this into four specialized modules: Planner, Executor, Verifier, and Generator, all coordinating through an evolving memory. 🔹 The "Secret Sauce": Flow-GRPO: The real innovation is in the training. Traditional offline Supervised Fine-Tuning (SFT) often leads to performance collapse in these systems. They introduce Flow-GRPO (Flow-based Group Refined Policy Optimization), a method that trains the "Planner" agent inside the multi-turn loop. It solves the sparse reward problem by broadcasting the final outcome reward back to every step of the reasoning chain. 🔹 The Results are Wild: By optimizing the planner "in-the-flow," their 7B model achieved massive gains over state-of-the-art baselines: +14.9% in Search tasks +14.0% in Agentic reasoning +14.5% in Math Crucially: It outperformed the ~200B parameter GPT-4o across all these domains. This is a must-watch if you are building AI agents or interested in how we can make smaller models significantly smarter through better system architecture and training. 📺 Watch the full breakdown here: https://lnkd.in/gSfPHbwh Github : https://lnkd.in/g4RaEcFg #AI #AgenticAI #LLM #MachineLearning #AgentFlow #OpenSource

25

How do we effectively combine retrieval and recommendation into a single pipeline? In my latest post, I explore the design space that sits between large-scale candidate generation and fine-grained ranking - an area often overlooked when we talk about these systems separately. 🔍 Some key points I dive into: • Why most real-world architectures split into a fast retrieval stage (ANN, inverted indexes, embeddings) followed by a personalized reranker • Techniques like joint training (e.g. multi-task or two-tower setups) that help the retrieval stage learn to surface better candidates for downstream ranking • Practical trade-offs: efficiency vs. accuracy, data requirements for training retrieval-aware rerankers, and deployment pitfalls like keeping embeddings up-to-date Real-world use cases from Netflix and research from Spotify shows significant gains in adopting this framework. If you work with large-scale recommenders, vector search, or end-to-end ranking, this one might give you some ideas - or at least a few design decisions to revisit. 👉 Check it out here: https://lnkd.in/gi3Z6Mvc Would love to hear your thoughts and feedback! 💬 #MachineLearning #RecommendationSystems #InformationRetrieval #AI #ML

162

10 Comments

Why mid-market Data Science teams can’t keep up Mid-market companies don’t have a data problem. They have a decision throughput problem. We see this pattern over and over again: • 2–10 data scientists • A backlog measured in months, not weeks • Teams stuck in analytics firefighting • Execs asking “what would happen if…?” • Answers arriving too late to matter The bottleneck isn’t talent or tools. It’s that Data Science has become the choke point for every important decision. Here’s why it breaks down: 1. Data Science capacity is scarce - Every new initiative competes for the same few people. - Most DS time gets burned on ad-hoc analysis, one-off models, and stakeholder explanations. 2. Time-to-impact is too slow - By the time an analysis is done, the decision window has passed. - Accuracy improves. Business outcomes don’t. 3. Talent retention is fragile - When decision logic lives in notebooks and individuals, the organisation is one resignation away from losing critical knowledge. 4. Executives need causal answers, not forecasts - “Will revenue go up?” isn’t the question. - “What happens if we change pricing, promos, or capacity?” is. Most teams can’t answer that fast enough. The uncomfortable truth: 👉 Hiring more data scientists doesn’t fix this. 👉 Better dashboards don’t fix this. 👉 More advanced ML doesn’t fix this. The constraint is decision velocity. The teams that break through don’t try to make Data Science faster. They make decision-making scalable, by turning bespoke analysis into repeatable, decision-grade systems that more people can use. That’s the shift from analytics as a service… to analytics as a decision engine. Curious how other mid-market teams are tackling this. What’s your biggest decision bottleneck right now?

6

4 Comments

95% of enterprise AI agent pilots will fail this year. Yep many of such headlines are already flooded in market. The reason? Most of the time we are treating multi-agent ecosystems like slightly smarter chatbots and context graphs are again just a band-aid that makes agent lives better but can't solve for architectural gaps. - Throwing a massive Context Graph at a problem won’t magically unlock autonomous decision-making. - Writing a better prompt won't fix a broken architecture. The Prompt Engineer role should evolve—we are now in the era of the AI Orchestrator/Architect. Why? When you scale from a single LLM to a multi-agent system, integration complexity explodes. You stop dealing with simple text outputs and suddenly have to architect: - State management across specialized agents - Hard decision boundaries and confidence scoring - Orchestration logic for "Experts as a Service" (EaaS) (many skip this step!!) The real innovation isn't coming from top-down R&D trying to build a monolithic god-model (and that is why I personally feel building new mid models froms scratch, while they are great, it just becomes chest thumping and will sit in fringe small usecases , not truely generalizable). It's coming from domain experts hacking together specialized agents to force cohesiveness into fragmented workflows. We are moving from AI as a standalone tool to AI as a digital assembly line - more standardization, better control on leakages and true process stabilization #assemblyLinesAreAwesome

20

2 Comments

People often ask me: “What does it actually take to grow into Product Data Science?” It’s one of the most exciting, fast-growing roles today. You sit at the intersection of product thinking, experimentation, causal inference, and applied AI, and you directly influence how product decisions get made. The truth is that A/B testing remains the foundation. And the complexity of running trustworthy experiments is only increasing. If you’re passionate about this, it’s a great moment to invest in your skills. Both of the A/B Testing courses below are currently 25% off: 1) A/B Testing 101 For PMs and Data Scientists beginning their experimentation journey • Fundamentals of A/B tests and how they drive product decisions • Writing hypotheses tied to business outcomes • Metric foundations, experiment setup and interpretation Prereqs: curiosity and a product mindset 🔗 https://lnkd.in/e4znBEqu 2) Advanced A/B Testing For Data Scientists who want to go deeper into experimental rigor • Handling pitfalls like SRM, peeking and noisy metrics • Sequential testing, variance reduction and Bayesian approaches • Making real-world ship or no-ship decisions 🔗 https://lnkd.in/eNHpn7mt These courses are intentionally designed to be sequential. → 101 builds the fundamentals. → Advanced takes you into high-rigor experimentation and decision science. 🎉 Both are 25% off this week. Don’t miss it.

27