⚡ Swift-native agents, micro:bit MPC and Infinite Brainrot -- top demos this week

Nat Taylor presented DSPy 101, a hands-on demo showing how to write AI programs with DSPy rather than prompts. He walks through a classification task: loading data, evaluating a base classifier, then optimizing and re-evaluating. DSPy keeps business logic separate from prompts and uses Python tooling for reproducible experiments. GitHub resources let builders inspect and reuse the approach for side-by-side comparisons. The takeaway: measurable experimentation with LLM settings resonates with builders and is practical for real-world pipelines.

Louis Knight-Webb from bloop presented How to stop being Claude's cuck, sharing Vibe Kanban for parallelizing Claude Code and other agents. Vibe Kanban is a Rust/TypeScript Kanban board that orchestrates AI coding agents and their CLIs, coordinating tasks across CI/CD pipelines. It offers practical orchestration patterns and a lightweight, on-device feel. Audience feedback hinted at solid interest in open-source, reproducible setups. It points to a path for developers to turn background agents into scalable AI tooling.

Fredy Alexander Rivera Acevedo presented Aquiles-RAG, a high-performance production-ready RAG that provides a lightweight abstraction layer to drop retrieval-augmented generation into services with minimal friction. The demo centers on Redis vector search with FastAPI and is embedding-agnostic, letting teams plug their preferred embedding model. Open-source and deployable by default, it aligned with the audience's preference for practical, production-focused RAG. It shows a practical path from prototype to scalable, real-world use for teams.

Guenther Lomas, founder & CEO of Sigtica, presented From Boilerplate to Boardroom: Unlocking Complex Legal & Financial Docs with PaddlePaddle. His talk walks through a parsing pipeline using Baidu PP-DocLayout-L for multi-column layouts, with Python handling footnotes. It compares PaddleOCR and Tesseract on identical texts, illustrating where each excels. LLMs like Baidu ERNIE structure outputs into clean JSON, in a containerized workflow. The demo offers a reproducible blueprint for production-grade document pipelines, and survey feedback hinted at strong practical value.

Isham Rashik from Future OS presented the YouTube Video Analysis Suite, a modular AI toolkit that extracts transcripts, generates knowledge graphs, performs sentiment and topic analysis, and supports note creation via natural language chat and a Streamlit dashboard. Built on FastMCP as the MCP server backend, with mcp-use as the client library, it demonstrates the power and flexibility of the Model Context Protocol (MCP) beyond frontend applications. The workflow, shown using Windsurf IDE and deployed with Docker, orchestrates multiple tools through LangChain and OpenAI models. The demo highlighted practical use cases for education and research, with potential to evolve into a broader copilots platform. Takeaway: multimodal analysis is now accessible for builders.

Thorsten Schaeff from ElevenLabs presented Making an ESP32 conversational with WebRTC. The setup pairs a tiny ESP32 (M5Stack AtomS3R) with an ElevenLabs voice clone, Pipecat AI, and WebRTC for real-time, low-latency audio. It showcases hardware constraints, why LLMs aren’t magic, and how open-source Pipecat contributions make the demo feasible. (People loved it) and it hints at practical edge AI for small devices.

Alexander Payne of Signal Eight presented From Chat to Action: Building Reliable AI Agents with Tool Orchestration, a demo of an AI-native Sales CRM that plans and orchestrates tasks. It defines tool schemas, manages workflows, and handles edge cases like ambiguity and idempotency, with a conversational UI that updates databases and logs interactions. The talk highlights production-ready agent patterns, showing how this approach scales to broader autonomous workflows beyond CRM. Takeaway: orchestration is becoming a practical path for autonomous AI.

Dan Hermes presented Writing Books with CrewAI, showing how autonomous AI agents coordinate to draft outlines, chapters, and a finished PDF from idea to publication. The demo presents CrewAI’s orchestration, Python scripts, and binding tools that drive a production-like workflow to generate machine drafts for human editorial pass. (audience feedback was positive) Hermes, author of Prompt Power and AI consultant, brings real-world depth from Lexicon Systems. It signals a practical path for scalable, writer-focused AI pipelines.

Bharat from DeepWeaver.ai presented Beyond One-Size-Fits-All: Building Intelligent LLM Selection Systems, a live demo of NVIDIA's LLM Router Blueprint. It presents a three-part system (Router Controller, Router Server with Triton, and downstream LLMs) deployed with Docker Compose and Kubernetes. The run maps code generation to DeepSeek, general Q&A to Llama 70B, and simple rewrites to Llama 8B, using Rust with Triton. Bharat, ML engineer and AI master's student, shows production-ready routing that teams can adapt for cost and latency awareness.

Jianzhun Du, co-founder of Splicing AI, presented Teamate, a proactive AI teammate inside Slack that listens to group conversations and steps in with suggestions and actions. Teamate lives in Slack channels and orchestrates with 100-plus integrations, shown via a live workspace and public invite. The demo walked through development workflows and real-time human AI collaboration. Subtle survey notes highlighted the practical interactivity as a strength. It hints at strong business potential if packaged as an on-demand teammate for distributed teams.

Yu Ching Lok, aka Alvin Yu, from Votee AI presented CantoneseLLM, a Cantonese-focused LLM designed to handle Hong Kong slang and local linguistic nuance. The demo outlined a data pipeline with hon9kon9ize community curation and low-resource language fine-tuning, plus benchmarking against HK Canto Eval. CantoneseLLM runs through Votee's enterprise-oriented platform and could enable localized AI tooling for Hong Kong businesses. Feedback noted practical value, while attendees seemed eager for more deployment details. It spotlights community-driven data curation as a path for localized AI.

Tabish Bidiwale from Fission presented OpenSpec, an open-source, spec-driven development framework that treats specs as code and integrates with any AI coding tool. Specs become central, testable artifacts that guide design and integration, distinguishing OpenSpec from tools like Kiro.dev. Built primarily in TypeScript and demonstrated with Claude and Cursor, it emphasizes reproducible workflows and safer shipping. The demo drew interest for its clarity and practicality, pointing toward a trend in production-ready, spec-driven AI tooling that builders can adapt to their pipelines.

Jamilton Alonso Quintero Osorio, Apiux Tech Lead, presented Agentes de voz que modifican tu interface en tiempo real. He demonstrated a real-time multi-agent voice system that interviews users by voice, using the Google ADK and Gemini Live API with low latency. The frontend uses Lit and TypeScript, with function calling tying intents to UI actions like showing cards and initiating contact, and the runtime runs on the GenAI SDK. Audience feedback was strong (people loved it).