⚡Local AI: Browser Agents + Ultra Low Latency Meeting Assist

⚡Local AI: Browser Agents + Ultra Low Latency Meeting Assist

Issue #3 · Week of August 25

Joe Heitzeberg • Founder at AI Tinkerers • ⏱️ 1 min read

Creating space for leading builders to share ideas, grow, and make an impact.

We pulled the standouts from AI Tinkerers meetups around the globe from the last two weeks: lots of agent orchestration and programmatic prompt optimization, with nods to on-device agents matching the recent push toward efficient, tool-savvy systems. Henry Mao (Singapore) shows MCP-first integration in Tool Calls are the New Clicks; Mahir Isikli (Berlin) compiles prompts with DSPy, not vibes; Rakesh Kumar (Chicago) runs private agents in the browser and Enrico Foschi showed how to achieve ultra low latency with locally-run voice apps. We scored for tech merit and survey feedback—real, vetted, practical builds. Read on.

Smithery: Tool Calls Henry Mao Co-founder / CEO at Smithery 📍 AI Tinkerers - Singapore • Aug 12 The future of the internet will be dominated by tool calls, not clicks. We're building Smithery to orchestrate the new era of AI-native services for AI agents. Henry Mao's demo showcased an AI agent leveraging multiple tools to perform a task autonomously and discussed the rising ecosystem of AI-native services (MCPs), and explained the technical underpinnings and approaches for enabling search and routing across MCPs to support agent tool discovery. Read more → TECH STACK GPT-4Claude-3Llama-2LangChainOpenAI API PROJECT LINKS smithery.ai Video @calclavia

DSPy Live Prompt Optimization Mahir Isikli Product Engineer at Arbio 📍 AI Tinkerers - Berlin • Aug 13 It was a demo of turning ad-hoc prompts into a repeatable optimization flow with DSPy. Mahir Isikli, an ML engineer, started with a brittle LLM pipeline, showed baseline accuracy, latency, and cost, then defined Signatures and Modules and ran an optimizer to generate candidates and re-tune prompts. He walked through the code and dataset wiring from the repo. Attendees left with a copyable recipe for prompt tuning; survey feedback hinted at its practical value for builders in our network. Read more → TECH STACK DSPyPythonJupyter PROJECT LINKS github.com dspy.ai @slobkebap

AWS: AgentCore & Kiro Pandurang Nayak ASEAN Head of Startups Solutions Architecture at AWS 📍 AI Tinkerers - Singapore • Aug 12 Pandurang Nayak pitched a demo on Recent Announcements from AWS, showcasing Kiro (an agentic coding IDE), AgentCore (new AWS agentic services), and Strands SDK for building agents. It highlighted cloud-native, production-friendly tooling that lowers barriers for agentic LLM apps, with practical demos and accessible repos. Pandurang, a seasoned AWS leader for ASEAN startups, conveyed solid execution. Feedback from the community hinted (people loved it), and the work points to a clear path for startup-scale automation and dev-prod workflows. Read more → TECH STACK Amazon BedrockStrands SDKKiro PROJECT LINKS amazon.com kiro.dev Video ArXiv Paper @pandurang

VeoSpark: Veo 3 Ad Pipeline Lewis T Founder at Foundry 📍 AI Tinkerers - Singapore • Aug 12 Lewis (lekt9), an AI engineer and open-source contributor with projects like alBERT-launcher, demoed VeoSpark, a conversational agent for AI video creation that reimagines workflows with tools like Veo 3 and Kling. Instead of prompt engineering, creators direct videos through natural conversation—starting with a script, then refining details like “make the boss more confused in scene 2”—while the system preserves what works. Under the hood, entity-based JSON outputs enable partial regeneration, context preservation, and intelligent stitching, reducing fatigue from repeated full regenerations. Early user tests showed high activation and iteration, and the demo drew strong enthusiasm, pointing to a scalable product for teams that want to direct videos like stories, not prompts. See also: videothon.getvideos.app - a competition for making the best video - credits provided. Read more → TECH STACK GPT-4Llama 2LangChainPyTorchOpenAI API PROJECT LINKS veospark.com tell.video Video @lekt8_

ClueWing Enrico Foschi Founder at Invisible Hours 📍 AI Tinkerers - Berlin • Aug 13 Enrico presented his side project, ClueWing, a fully local meeting assistant, using local c++ modules for transcription (whisper.cpp - 30ms latency), llm processing (llama.cpp - 250ms processing on macbook m1 air 8gb), and a Flutter app able to get an overall realtime meeting coaching. All at 20% cpu usage. Read more → TECH STACK GPT-4Llama-2Claude-3LangChainOpenAI API PROJECT LINKS cluewing.com

More Great Builds

Quick hits from the community — demos worth bookmarking:

Gemini YAML Generator

Paolo Perazzo • AI Tinkerers - San Francisco • Aug 14

Paolo Perazzo, head of AI at Axiom and a seasoned AI product executive, presented a production-focused AI assistant that generated Kubernetes and Sveltos YAML from natural language, using Gemini 2.5 Flash and a multi-stage agent workflow that classifies intent and routes to specialized agents. He demonstrated function calling to enforce complex, nested Kubernetes schemas and a RAG pipeline with Google's embeddings and Search Grounding. The project on GitHub at https://github.com/sivola/Sveltos-Assistant showed code and an actionable pattern for enterprise config automation.

Loading tech tags...

github.comsubstack.comVideoArXiv Paper

Compilers of Intent

Patrick Boueri • AI Tinkerers - Chicago • Aug 14

Compilers of Intent explores how LLMs change how we communicate with our computers. The experimental Intentc implements a build framework for compiling our written intent into working code, similar to GNU Make. Concepts explored include validation, refinement, reproducibility, and how we may express our intent more cleanly. Patrick Boueri, Director of Data Science at Indeed, delivered a hands-on demo and highlighted the open-source repo for builders to inspect. The session drew positive feedback for its nuts-and-bolts approach and he solicited collaboration to make intentc a real product.

Loading tech tags...

pboueri.github.iogithub.com

Chatbot UI Control

Andrej Böttcher • AI Tinkerers - Hamburg • Aug 14

It showcased 'Beyond the Prompt: Building Chatbots That Control Your UI,' a demo where a chatbot directly manipulates a frontend and presents interactive in-chat controls. It used LLM tooling to render UI components and commands, with automatic page context giving the bot awareness. The presenter, a Hamburg-based principal engineer with a GitHub track record, impressed with the practical wiring between chat, UI, and APIs. Audience reactions were positive to hands-on approach, and it hinted at business potential for automated UI workflows.

Loading tech tags...

Browser Private Agents

Rakesh Kumar • AI Tinkerers - Chicago • Aug 14

Rakesh Kumar presented 'Private Agents in your browser,' a quick demo showing small language models running agent tasks entirely in a laptop browser. It highlighted privacy-first, browser-based agent workflows that cut latency and cloud costs, with a live demo and GitHub resources. Rakesh, founder of Techne, brings deep AI engineering know-how. The audience reacted positively (people loved it): the browser-on-device approach fits current trends and offers actionable patterns for startups and on-device products.

Loading tech tags...

techne.appVideo

Codapt: AI Planning

Trevor Keith • AI Tinkerers - San Francisco • Aug 14

Trevor Keith, a technical founder at Codapt, walked through how to maximize planning power for AI agents, drawing on ~16 iterations to craft production-grade Next.js apps. The demo covered planning pipelines, context population, and subtasking to balance latency and code quality, with a public GitHub repo for hands-on exploration. It felt grounded in real feedback (people loved it) and practical patterns for scalable agent workflows. If turned into a product, these pipelines could speed up AI-driven dev tools for teams.

Loading tech tags...

codapt.ai

You are one of 95,000+ readers from Anthropic, OpenAI, Google, Microsoft, Meta, Apple, Amazon, Nvidia, Netflix, Stripe, Databricks, Snowflake, and others — spanning frontier labs, big tech, startups, and top universities.