LAI #77: Structured Outputs, LangGraph NLP, Sub-ms Agents, and Personalization at Scale
Offline resume tools, startup collabs, and the hidden infrastructure powering real-time AI systems.
Good morning, AI enthusiasts,
This week’s issue is a mix of applied AI and infrastructure that actually scales. We start with a deep dive into structured output from both local and cloud-based LLMs — Gemini, Gemma, and SmolLM all in the mix. From there, we explore graph-based NLP with LangGraph and Large Concept Models, a new angle on sentiment analysis and idea-level processing.
We’re also looking at how MCP and Google’s ADK enable sub-millisecond agent communication, plus a technical guide to fine-tuning DeepSeek-VL2 for multimodal instruction following. And in What’s AI, we zoom out to something more familiar: the Netflix thumbnail, why yours doesn’t look like your friend’s — and what that says about how personalization with AI actually works.
Let’s dive in.
What’s AI Weekly
This week in What’s AI, I will explore something fun: why the old one-size-fits-all strategy in ads and marketing is obsolete, and how AI is revolutionizing marketing by making it personal and engaging. Try this: open your Netflix account and search for a movie, and ask a friend to do the same on their account. Notice anything different? The thumbnails you both see are probably not the same, even for the same shows! Netflix customizes them just for you, based on what you’ve watched and liked. This is all thanks to AI using techniques, and that’s exactly what I am diving into this week. Read the complete article here or watch the video on YouTube.
— Louis-François Bouchard, Towards AI Co-founder & Head of Community
Learn AI Together Community Section!
Featured Community post from the Discord
Typesafeui has built a resume builder with features like an ATS-friendly template, version history, and drag and drop to reorder sections. It is mobile-friendly and works entirely offline. It has a clean UI, check it out here, and support a fellow community member. If you have any feedback or suggestions, share them in the thread!
AI poll of the week!
57% of you believe no industry will remain untouched by AI, and I agree; it might just take longer for some sectors to adopt AI. But what about highly human labour industries where humans might still be cheaper to hire than building and maintaining AI systems? Share your thoughts in the thread!
Collaboration Opportunities
The Learn AI Together Discord community is flooding with collaboration opportunities. If you are excited to dive into applied AI, want a study partner, or even want to find a partner for your passion project, join the collaboration channel! Keep an eye on this section, too — we share cool opportunities every week!
1. Efficientnet_99825 is writing a research paper on recommendation systems and methods, along with improvements to date. He wants to create code bases for it, too. He is looking for someone experienced to guide or collaborate on the paper. If this sounds like you, reach out to him in the thread!
2. Sweatysteve123 is looking for an AI/ML partner to build a startup. If you are in the Bay Area or around and would like to explore the idea, connect with him in the thread!
Meme of the week!
Meme shared by ghost_in_the_machine
TAI Curated section
Article of the week
Structured Output With Local and Cloud-Based LLMs By Robert Martin-Short
This article details a case study on extracting structured recipe data from images using cloud-based (Gemini 2.5 Flash) and entirely local LLM pipelines (Gemma3 and SmolLM2–1.7B with the outlines package). It outlines methods for achieving schema-compliant output, including constrained decoding for local models. The piece also touches on using the Google Drive API for data management and LLM-Judge (Claude 3.7) for comparative evaluation, finding cloud models currently superior but noting rapid advancements in local capabilities.
Our must-read articles
1. Graph-Based NLP with LangGraph and Large Concept Models(LCMs): Sentiment Analysis and Beyond By Samvardhan Singh
The author explored graph-based Natural Language Processing, focusing on Large Concept Models (LCMs) that process entire ideas rather than individual words, enhancing semantic understanding. It detailed how integrating LCMs with graph structures, managed by frameworks like LangGraph, improves contextual reasoning and relationship mapping within complex text data. A hybrid symbolic-semantic model and a practical LangGraph pipeline for sentiment analysis of customer feedback demonstrated the system’s ability to extract nuanced insights.
2. The Hidden Power of MCP + Google ADK — A Guide to Building Systems That Scale By Subhadip Saha
The author discussed Model Context Protocol (MCP) servers and Google’s Agent Development Kit (ADK) as solutions for enabling AI models to interact with external services. MCP servers standardize AI communication with tools like APIs and databases, enhancing security and flexibility. ADK provides a framework to build and deploy AI agents utilizing these connections, resulting in more capable and scalable systems. The discussion covered integration methods, practical applications, and key considerations for security and performance.
3. Fine-Tuning DeepSeek-VL2 for Multimodal Instruction Following: A Comprehensive Technical Guide By Ojasva Goyal
This technical guide explained the fine-tuning of DeepSeek-VL2, a multimodal model, for Visual Question Answering. It detailed dataset preparation, adherence to the model’s chat template, and using LoRA for efficient adaptation. Solutions to technical challenges like tensor mismatches and compatibility issues were presented, which resulted in reduced VRAM usage and improved task accuracy.
4. This MCP + ADK Combo Is My Secret Weapon for Sub-Millisecond Data Streaming By Subhadip Saha
The article explained how the Model Context Protocol (MCP) and Agent Development Kit (ADK) enable sub-millisecond data streaming for agent-to-agent communication. It detailed MCP’s role as a lightweight streaming layer and ADK’s function in building efficient AI agents with optimized data serialization. It emphasizes the significance of low latency in fields like autonomous systems and high-frequency trading and compares this approach to traditional methods.
If you want to publish with Towards AI, check our guidelines and sign up. We will publish your work to our network if it meets our editorial policies and standards.
https://substack.com/@cortexmuteek?r=5re9la&utm_medium=ios
This is worth checking out