AI Tinkerers #18: Community Spotlights • MemoryLane Screen Context & On-Device Voice Models
AI Tinkerers #18: Community Spotlights • MemoryLane Screen Context & On-Device Voice Models
Issue #18 · Week of March 16
Joe Heitzeberg • Founder at AI Tinkerers • ⏱️ 1 min read
Creating space for leading builders to share ideas, grow, and make an impact.
This week’s AI Tinkerers demos were all about what happens above the model layer: how agents remember, coordinate, and where they’re allowed to act autonomously.
Multi-agent orchestration is the biggest theme by far.Skwad from Nicolas Bonamy hosts multiple coding agents in a unified IDE, wayyOS from Rick Galbo reimagines a desktop OS as a swarm manager, and Alex Lebedev showed how PostHog pipelines detect issues and route them to coding agents that ship PRs. The question has shifted from “can an agent code?” to “how do I run a team of them?”
Context and memory keep coming up as the bottleneck builders want to solve. Filip Kubiš in Prague demoed MemoryLane, which watches your screen and pipes that context into AI tools through MCP so you stop re-explaining yourself. Vineeth Voruganti showed Honcho, which gives agents long-term memory so they self-improve across sessions. Both are attacking the same problem: every new conversation starts from zero, and it shouldn’t.
Voice-first and on-device AI showed up in force.UnaMentis from Richard Amerman is pushing on-device voice models for the lowest possible latency, Sebastián Franco Gómez demoed using your phone as a voice-controlled agent interface that executes desktop tasks remotely, and Story Jam from claro fix turns spoken words into music in real time.
Agents are pushing into regulated and real-world domains. Simon Redfern demoed Opey executing payments in banking APIs via delegated consent, Andy Singleton showed AI owning software IP through tokenized licenses, and Tinyfish from Trần Thị Thắm replaces brittle XPath scraping with agents that just navigate pages given a goal.
Congrats to everyone featured this week. Read on for details!
Nicolas Bonamy from Kochava presented Skwad, a multi-agent IDE, and explained why he built it. Skwad is a macOS app that orchestrates parallel AI agents, coordinated through libghostty and MCP. The approach keeps agent collaboration lightweight and accessible, with clean orchestration and GitHub integration, and survey feedback hinted strong interest. It shows a practical path toward desktop multi-agent tooling that could accelerate real world development.
Andy Singleton from Maxos presented an AI that launches, runs, and owns a software company, showcasing a Tokenized IP License that lets AI own software and IP by tokenizing GitHub repos with TIPL. The demo shows a conversation with OpenClaw to research, launch, promote, and build a project, illustrating a loop of AI-enabled operations and governance. It hints at scalable business models, a theme survey feedback found timely. For builders, it highlights practical paths to autonomous product creation.
David Zhang presented An engineering team in your pocket, a meta-harness that coordinates an agentic engineering team on the go. The demo uses a Python-based state machine to orchestrate asynchronous coding agents via markdown tickets, with open-source repos and documentation backing it up. Survey vibes hinted at strong interest in its production-grade agent loops that auto-write, test, and verify code, and for the pocket-team spirit that scales long-horizon projects. It shows orchestration is moving toward practical, autonomous tooling for builders.
Richard E Amerman from UnaMentis showcased UnaMentis, a mobile AI voice-first learning platform. The demo highlighted ultra-low latency, on-device TTS running directly on iOS via a Rust/Candle port, and a provider-agnostic, modular edge stack built around Kyutai Pocket TTS UMCF. It's pivot-ready and open, with open-source resources for porting advanced models to edge devices. This approach points to practical, real-time voice tutoring on-device, a strong fit for mobile, latency-sensitive AI apps.
Rick Galbo from Wayy Research presented wayyOS, an agentic swarm manager for a desktop environment. wayyOS orchestrates a desktop swarm with global context and inter-AI messaging, built on an OpenClaw-inspired harness. It runs as the workstation, handling daily tasks and automations. Survey feedback hinted strong interest in practical, deployable agentic workflows. If released as a product, it could unlock real-time, self-managing desktop automation for power users. It shows how tight orchestration and lightweight agents move from demo to everyday tooling.
Claro fix presented Speech to Music with Story Jam, a live demo that turns audience words into chord stories via Elevenlabs transcription and a custom music-generation model. The setup runs in near real time, with transcription feeding a generative audio system that outputs evolving melodies on the fly. It showcases multimodal pipelines, low-latency orchestration, and a path toward interactive soundtrack tools for narrative media. It matters for builders as production-ready audio AI that adapts to spoken narratives.
Craig Robinson from Playco presented Clawber.ai, an Agentic Battle Arena where AI agents write JavaScript to drive bots in a live ladder. Clawber.ai is an agent-first platform with an onboarding funnel, API self-registration, sandboxed code running on AWS Lambda, and 24/7 matches with live spectating and a public leaderboard. Humans verify ownership via Twitter, turning signups into public buzz. (audience noted its novelty) This demo hints at scalable agent ecosystems and real-time, edge-friendly onramps for AI agents.
Trần Thị Thắm, a TinyFish ML Gen Engineer, presented Real-Time Web Scraping with AI Agents: No Puppeteer, No XPath, Just a Goal. The demo shows AI agents autonomously navigating Booking.com and Agoda to return structured JSON, with a dashboard. It runs a Next.js app built in TypeScript, Tailwind CSS, and PostCSS, and demonstrates goal-driven browsing rather than brittle DOM scraping. Open-source code is provided for builders, and audience feedback noted its practicality. It hints at production-ready data collection in apps.
Sean Kelly from OCLC presented Local LLMs, Free Guy and Text Adventure Games, a Cursor web editor for building AI-generated worlds and objects. The demo runs on local LLMs and uses semantic evaluation to keep world rules flexible, with edits playable in time. It felt impressive (people loved it) to see a storytelling loop that runs without cloud dependencies. For builders, it hints at practical tools for AI-driven game design. Takeaway: local, composable AI can power flexible, production-ready creative workflows.
Ky-Nam Nguyen from EngageKit presented Streaming Dynamic Selection for Contextual Engagement on Social Media. His Chrome extension EngageKit.io uses context-aware AI to dynamically select persona for human-in-the-loop LinkedIn engagement in real time. The team shared open-source access on GitHub, showcasing a streaming pipeline that keeps persona management reliable across long-horizon interactions. Audience feedback was positive, with a quiet nod to its real-time persona control. It’s a practical blueprint for agentic social tools that could scale into production-ready outreach platforms.
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.
🛠️
Want to be in the next issue?
Every demo here came from someone who showed up and built something.
![AI Tinkerers #18: Community Spotlights • MemoryLane Screen Context & On-Device Voice Models [AI Tinkerers - Post-Training]](https://sloppy-joe-app.imgix.net/ai_tinkerers/logos/post-training_laptop-sticker_190_de2487c31e50813a3471e79627268ffb.png?w=300&fit=max&auto=format%2Ccompress&q=75){kind=small}
AI Tinkerers #18: Community Spotlights • MemoryLane Screen Context & On-Device Voice Models