Rahil Bagheri

Really proud of the team on this one. Building our own Deep Research methodology here at Publicis Sapient has allowed us to tailor how enterprises run deep research—so they can generate high-quality, grounded, and truly competitive intelligence, not just generic AI outputs. What’s exciting is how this is evolving. We’re now combining Bodhi Deep Research with our Agentic Platform to build industry- and business-specific offerings that enterprises can actually deploy and scale. And the best part — these capabilities work across clouds as well as on-prem environments, making them practical for real-world enterprise setups. We strongly believe Agentic transformation will be a key differentiator for organizations in the coming years. The ability to orchestrate AI agents, research workflows, and decision intelligence will define how companies compete. Great to see the Bodhi ecosystem ( https://lnkd.in/gBja-FNT ) being battle-tested in these scenarios and delivering real value. Excited for what’s coming next. #SapientBodhi #AgenticAI #DBT #AITransformation #EnterpriseAI

128

5 Comments

In the era of LLMs, evaluation isn’t optional - it’s foundational. Came across DeepEval, an open-source framework purpose-built for testing LLM outputs. Link in comments #LLMEvaluation #GenerativeAI #LLMs #DeepEval

64

1 Comment

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

AI may have hit “peak data”. Goldman Sachs’ data chief Neema Raphael has warned that large models like ChatGPT and Gemini have already run out of fresh human-made training material. From here on, much of the data feeding AI will be synthetic - generated by AI itself. What happens when machines are learning mainly from their own outputs? Do we risk a creative plateau? How do we keep models grounded in human context, nuance, and originality? Others in the field have drawn comparisons with fossil fuels: human-created content is finite, and we may already have burned through most of it. The internet is vast, but not infinite. Source - https://lnkd.in/eQHVN4xS

9

1 Comment

🚀 Just wrapped up a hands-on project applying Retrieval-Augmented Generation (RAG) to a real-world use case: interpreting LIC’s New Jeevan Shanti policy document using LLMs. The core idea? Combine vector search over regulatory PDFs with GPT-4 to generate grounded, faithful answers to complex policy questions. Key capabilities we built: ✅ LCEL-based RAG architecture using LangChain ✅ Contextual compression retriever for focused document lookup ✅ Modular LLM & embedding selection (OpenAI, HuggingFace, Mistral) ✅ Evaluation using LLM-powered metrics: faithfulness, relevance, correctness 📄 The document used is publicly available here: LIC New Jeevan Shanti Policy PDF 💻 Full code available on GitHub: https://lnkd.in/gS9mEzWA 💡 This architecture brings the best of both worlds — retrieval accuracy and generative flexibility — especially in regulated domains like insurance and financial services. 🔜 Next step: Incorporate Agents that can reason across multiple steps — not just answer a question, but decide which tools to use, when to search, how to validate, and when to stop. That’s where the future of enterprise GenAI lies. Would love to hear how others are pushing RAG, agents, and LLMs in production! #GenAI #LangChain #LLM #RAG #AIForOps #InsuranceTech #LLMEvaluation #LangChainAgents

61

1 Comment

2025 is the Year of LCMs and not LLMs. Meta has announced a new architecture for the future of Large Language Models called Large Concept Models. Building upon the foundations laid by LLMs, LCMs leap forward in AI's ability to understand and generate human-like text. Here's how LCMs work: 1️⃣ Conceptual Processing LCMs encode sentences as unique "concepts," enabling high-level reasoning and contextual understanding. 2️⃣ SONAR Embeddings These embeddings capture the semantic essence of a sentence, transcending word-level processing. 3️⃣ Diffusion Techniques LCMs use diffusion methods to stabilize outputs, leading to consistent and reliable results. 4️⃣ Quantization Methods Quantization techniques enhance robustness and reduce errors from minor perturbations. 5️⃣ Multimodal Integration LCMs support multiple modalities, including text and speech, facilitating cross-lingual comprehension. Whether you're generating detailed reports or engaging in complex reasoning, LCMs elevate AI capabilities, offering deeper understanding and more structured outputs.​ Here's how LCMs are architecturally different from LLMs: LLMs: - Operate at the token level, predicting the next word or subword in a sequence. - Utilize transformer-based architectures for sequential token prediction. - May struggle with maintaining long-range coherence in text generation. LCMs: - Process input at the sentence or concept level, capturing broader semantic meaning. - Use SONAR embeddings to map sentences into a language-agnostic semantic space. - Excel in hierarchical reasoning and abstraction, enabling high-level reasoning and contextual understanding. Understanding these distinctions is essential for selecting the appropriate model for specific applications, ensuring more effective and contextually appropriate AI interactions. LCMs aren't just more advanced; they're more insightful: ✅ Process entire sentences as unified concepts. ✅ Grasp context and nuance beyond individual words. ✅ Generate coherent and contextually relevant outputs. LCMs is essential. They save time, boost productivity, and create a more natural flow in AI-human interactions. Over to you: What tasks do you think would benefit the most for LCMs? Image credits: Manthan Patel

324

83 Comments

AGI will not come from scaling alone. Six serious pathways beyond today’s LLM paradigm (and what data scientists should watch). The Computing Community Consortium’s new report sketches six plausible research tracks that could define the next AI era. These are not replacements for LLMs but directions that could expand or even reroute them. 1. Neuro-Symbolic AI Combining neural learners with symbolic logic for better reasoning and control. Watch: benchmarks that test compositionality and explainability, and open-source libraries fusing logic with gradients. 2. Neuromorphic AI Brain-inspired hardware for faster, more efficient edge computing. Watch: systems like IBM NorthPole and Intel’s Loihi 2 that are starting to move from labs to real workloads. 3. Embodied AI AI that acts in the physical world to gain grounded knowledge and causal understanding. Watch: humanoid robots like Apptronik and Figure integrating real-world perception with AI reasoning. 4. Multi-Agent (Agentic) AI Networks of cooperating or competing AI agents that plan and execute tasks. Watch: frameworks like AutoGen and OpenAI Operator, and research testing when collaboration beats a strong single model. 5. Human-Centered AI AI that aligns with human values, ethics, and social intelligence. Watch: design patterns from human-AI collaboration research and regulatory frameworks like the EU AI Act. 6. Quantum AI Merging AI with quantum computing for massive speedups in search and optimization. Watch: error-correction milestones from Google and IBM’s roadmap toward fault-tolerant systems. What this means for data scientists • Structure your models. Add constraints and logic checks to improve reasoning. • Pilot agents responsibly. Use multi-agent frameworks where coordination adds value, but test against strong single models. • Prepare for regulation. Treat transparency and oversight as core design elements. • Track hardware. Follow neuromorphic and quantum roadmaps for future cost and performance inflection points. If you had to place a three-year bet, which two of these pathways do you think will matter most in your stack? If you found this useful, please consider reposting or tagging someone who’s thinking about where AI is headed next. I’d love to hear which of the six pathways you’re most excited about. #AI #AGI #DataScience #MachineLearning #NeuroSymbolic #Agents #Robotics #QuantumComputing #HumanCenteredAI #ArtificialIntelligence #AIResearch #FutureOfAI #MLOps #GenerativeAI #DeepLearning

5

🤖 How RAG systems are built -> Retrieval-Augmented Generation (RAG) is one of the best approaches you can take to reduce hallucinations in an LLM, and make it answer your questions based on data that was not part of the training dataset. My latest side project is a RAG that runs 100% in your computer, including the embedding generation and the LLM inference. No external APIs are used. In this article you will find a complete explanation of how it works, and the principles behind this powerful technique 👇

12

4 Comments