Introducing Agentic AI: A New Blog on Problem Formulation and LangChain

Could Agents make their own custom models on command? I was fascinated to read about “LoRA adapters On Tap” in the latest issue of The Batch from DeepLearning.AI by the ever brilliant Andrew Ng. It seems that researchers are now producing models that will create fine-tuning LoRA matrices based on a text description of the task required. This innovation will put an incredibly powerful customisation tool in the hands of AI Agents allowing them to fine-tune LLMs at will for tasks as they arise - talk about pulling yourself up by your bootstraps! Link to the article: https://lnkd.in/eYUV9gdZ

In the latest issue of The Batch, Andrew Ng announces his latest course, Agentic AI, a hands-on builder course using four key design patterns (reflection, tool use, planning, and multi-agent collaboration). Plus: 🗞️ Anthropic launches Claude Sonnet 4.5 and overhauls Claude Code 🗞️ OpenAI, Meta diversify AI product lines 🗞️ Alibaba adds Qwen3-Max and open multimodal Qwen3-VL/Omni models 🗞️ LoRA adapters on tap Read The Batch: https://hubs.la/Q03NnHFR0

🚀 Day 1 of My Gen AI Journey Begins The goal is simple: learn, build, and share every single day — from transformer fundamentals to deploying intelligent, production-ready AI agents. What I explored today — “How LLMs read text” Before you can build with large language models, you have to understand how they see language. So today was all about tokenization — the first and most crucial step in any LLM pipeline. what I worked on: 👇 1) Studied how text is converted into tokens using Byte-Pair Encoding (BPE) — merging character pairs based on frequency to form subword units. 2) Implemented my own BPE tokenizer in Python from scratch (reading corpus → counting pairs → merging → generating vocab). 3) Compared it against OpenAI’s tiktoken library to understand how GPT-style models chunk input text — and realized how even simple words can split into multiple tokens based on frequency and merges. 4) Visualized token-to-ID mapping to see how vocabulary evolution affects sequence length and compute cost. That hands-on experiment made me appreciate how much efficiency and compression go into even reading text before any model layer starts processing it. What I built: 1) tokenizer.py → my custom BPE implementation with encode/decode functions 2) micro_transformer.py → a 4-layer transformer using PyTorch, built around the core components: -Scaled dot-product attention -Multi-Head Attention -Position embeddings -LayerNorm + residuals 3)app.py → served the model using FastAPI, exposing a /predict endpoint to test next-token predictions in real time 🧩 Key takeaways 1) Tokenization alone changes sequence lengths, memory footprints, and attention complexity (O(n²) cost). 2) Vocabulary size is a trade-off: too small → more tokens per word; too large → higher embedding matrix size. 3) tiktoken is extremely optimized in C and Rust under the hood — understanding its internals helps you reason about latency and cost in real deployments. Tomorrow I’ll dive into embeddings — how tokens become high-dimensional vectors that capture context and meaning. We’ll explore cosine similarity, vector normalization, and embedding models like text-embedding-3-large and bge-m3. Question for you all: What part of LLM internals would you like me to explore next? #GenerativeAI #LLM #MachineLearning #AIEngineer #OpenAI #tiktoken #Transformer #PyTorch #FastAPI #DeepLearning #LearningInPublic #AICommunity

🚀 New Blog Live! Just published "10 Essential Key Components of RAG System Every AI Engineer Should Know" If you’re exploring AI and LLMs, understanding RAG (Retrieval-Augmented Generation) is a game-changer. This blog breaks down the key components of RAG, showing how it empowers LLMs with accuracy, scalability, and context. Perfect for AI engineers, data scientists, and tech enthusiasts looking to level up their AI projects. 👉 Check out the full blog here: https://lnkd.in/epd3vw3G #AI #MachineLearning #RAG #RetrievalAugmentedGeneration #LLM #ArtificialIntelligence #DataScience #TechBlog #AIEngineering #AIResearch

I just published an article that explains how AI agents work by connecting large language models with real-world tools. It includes a hands-on example of building a functional AI News Agent using Llama 3.1 and Ollama.

New tutorial! 🚀 KV Cache Optimization via Multi-Head Latent Attention 🧠 Cut KV cache bloat without losing accuracy—see how MLA shrinks memory and speeds up tokens. 🧩 Uses Multi-Head Latent Attention (inspired by DeepSeek) + RoPE, with tooling like PyTorch and vLLM. ⚙️ Built with hands-on PyTorch code, toy Transformer, and clear plots to compare latent dims. https://pyimg.co/bxvc0 👍 Author: Puneet Mangla #LLM #Inference #KVCache #MultiHeadLatentAttention #RoPE #DeepSeek #PyTorch #vLLM #AI #MLOps

🧠 Breakthrough in AI Reasoning: The Hierarchical Reasoning Model (HRM) Excited to share groundbreaking research from Sapient Intelligence that challenges how we think about AI reasoning! The Problem with Current AI 🤔 🎇Despite their impressive capabilities, Large Language Models face fundamental limitations: ✈️Chain-of-Thought (CoT) methods are brittle and data-hungry Fixed-depth architectures prevent true algorithmic reasoning State-of-the-art models struggle with complex tasks requiring deep reasoning Enter HRM: Brain-Inspired Architecture 🚀 Inspired by how the human brain organizes computation hierarchically across different timescales, researchers developed the Hierarchical Reasoning Model with just 27M parameters. Key innovations: Two interdependent recurrent modules: high-level for abstract planning, low-level for detailed computations Hierarchical convergence that avoids premature convergence issues in traditional RNNs One-step gradient approximation that eliminates memory-intensive BPTT Trained on only 1,000 examples—no pre-training required! Remarkable Results 📊 HRM achieves what seemed impossible: ✅ 40.3% on ARC-AGI-1 (vs 34.5% for o3-mini-high, 21.2% for Claude 3.7) ✅ Near-perfect accuracy on 9×9 Sudoku-Extreme (55% vs 0% for all baseline models) ✅ 74.5% on 30×30 Maze navigation (vs 0% for CoT-based approaches) All with 30× fewer parameters and 30× smaller context windows than competitors! Why This Matters 💡 HRM represents a paradigm shift from token-level reasoning to latent computational depth. It shows that: Hierarchical processing enables universal computation capabilities Small, specialized models can outperform massive general-purpose LLMs on complex reasoning Brain-inspired architectures offer a viable alternative to the scaling-focused approach The model even exhibits emergent properties matching neuroscientific observations about cortical hierarchy! Looking Ahead 🔮 This work opens exciting possibilities for: More efficient AI systems Better handling of algorithmic reasoning tasks Integration with existing architectures Congratulations to the team at Sapient Intelligence for this important contribution to AI reasoning! 📄 Read the full paper: https://lnkd.in/e3pftGnp 💻 Code: github.com/sapientinc/HRM #AI #MachineLearning #DeepLearning #Reasoning #Research #ArtificialIntelligence #NeuralNetworks #AGI

Turning RAG into an Information Sieve: A fresh take on Retrieval-Augmented Generation (RAG) I recently encountered "How to Turn RAG into an ‘Information Sieve" by Florian June, which offers an insightful exploration of evolving RAG from a basic retrieval method into a structured, intelligent reasoning pipeline. Key takeaways include: - Traditional RAG treats a user query as a single block: retrieve → synthesize → answer, often missing context and depth. - DeepSieve introduces a smarter approach by breaking queries into subproblems, routing them to the appropriate data sources, rerunning failed paths, and fusing the results into coherent, accurate answers. - The process (Decompose → Route → Reflexion → Fuse) enhances transparency, adaptability, and performance in real-world AI applications. - This model excels with heterogeneous data—text, APIs, structured databases by maintaining context across all layers. Why this matters: As we integrate GenAI systems into enterprise and multi-modal environments, intelligent retrieval and structured reasoning can significantly reduce hallucinations and improve output fidelity. I highly recommend giving it a read: How to Turn RAG into an “Information Sieve” #AI #MachineLearning #GenAI #LLM #RetrievalAugmentedGeneration #DeepSieve #ArtificialIntelligence #NLP #DataScience #Innovation #TechTrends #AITech #InformationRetrieval #RAG

🚀 From Text Chatbot → Voice-First AI Assistant It’s been a while! I've been deep in the build-and-learn phase, and I'm excited to break my silence with a project I'm truly proud of: RDL Voice AI Assistant — a real-time voice chatbot for RDL Technologies that delivers accurate, context-aware answers from a custom knowledge base. ⚡ What Makes It Special LiveKit – Real-time voice communication Google Gemini – The reasoning brain behind responses Deepgram & Cartesia – Speech-to-text & lifelike text-to-speech RAG (Retrieval-Augmented Generation) – AI grounded in RDL-specific data 🗣 How It Works 1️⃣ Speak into the mic 2️⃣ Deepgram transcribes your voice 3️⃣ Gemini checks if RDL knowledge is needed 4️⃣ RAG fetches relevant info from our FAISS database 5️⃣ Gemini generates an informed answer 6️⃣ Cartesia converts it to natural speech 7️⃣ You hear the AI respond With RAG → precise company/product insights Without RAG → generic LLM answers ⚡ Built for Real-Time Asynchronous Loading: Faster startup Connection Pooling: Efficient API usage Memory Management: Context kept smartly Error Handling: Stable even under load This was a deep dive into voice AI + retrieval-augmented intelligence, and I’m thrilled with the outcome. 📎 Check out the code & README: https://lnkd.in/gSFydejf The detailed README covers everything from environment setup to launching the agent. I’d love to hear your thoughts or suggestions! What’s the most interesting project you’ve worked on lately? #AI #VoiceAI #RAG #LLM #GenerativeAI #Python #DeveloperJourney #OpenSource #AgenticAI

The Size Myth is Busted: A 7M-Parameter Model Just Crushed LLMs on Logic! We've been told for years that the key to AGI is scale. But what if the real breakthrough is architecture, not size? Is this a game-changer for efficient AI, signaling a shift toward specialized, highly effective small models for complex problem-solving? Read the full breakdown of the TRM's simple yet powerful architecture and its mind-blowing results: https://lnkd.in/duc4tD_i #AI #MachineLearning #LLMs #DeepLearning #ResponsibleAI #Innovation #Logic