Top AI Demos #25: Voice Agents, MLX Trading Cards & Cyber Threat Intel

Top AI Demos #25: Voice Agents, MLX Trading Cards & Cyber Threat Intel

Issue #25 · Week of May 4

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

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

We’ve reached some kind of inflection point in the quality and quantity of demos this week. We’re seeing a strong trend towards agents that can reliably perform complex tasks, whether it’s Adrian Stobbe’s work on assessing AI voice-agents for leadership training or Alex Key’s approach to agentic 3D point cloud editing using LLM spatial reasoning.

Several builders are focusing on practical, local-first AI solutions. Felix Kemeth’s demo on fast trading card grading with MLX and visual chain-of-thought on consumer hardware, and Eric Fillion’s work on running large AI models offline on Macs with Eric Chat, highlight this shift.

We’re also seeing advancements in how AI systems maintain continuity and context. Vinay Guda’s Gemini Live voice agents demonstrate low-latency multimodal interactions, while Nanda Kumar’s ThreadHop explores local-first context sharing for coding agents.

Read more »

Sponsored by Coral

Open-source data retrieval layer for agents

Agents can make fewer, more precise queries with Coral. Turn any API, database or file to SQL tables. Benchmarked at 2x more token efficiency than MCP and CLI tools, with higher accuracy and security.

Assessing AI Voice-Agent Leadership Training Adrian Stobbe Software Engineer at Insiderisk 📍 AI Tinkerers - Valencia • Apr 21 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS insiderisk.com @astobbe_

Agentic 3D Point Cloud Editing Alex Key AI Engineer at panoriq.ai 📍 AI Tinkerers - Nürnberg • Apr 22 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS panoriq.com @alexkey_me

MLX: Fast Trading Card Grading Felix Kemeth Senior Data Scientist at Akribic 📍 AI Tinkerers - Nürnberg • Apr 22 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS akribic.com Video

AI Cyber Threat Intelligence Dashboard Vidyut Rajagopal Student at Georgia Institute of Technology 📍 AI Tinkerers - Atlanta • Apr 21 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS github.com

office-agents: Office AI agents Li Yang Hew AI Engineer at brightriver.ai 📍 AI Tinkerers - Singapore • Apr 21 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS github.com @hewliyang

Read more »

Sponsored by Kiro

New in Kiro CLI 2.0

Prompt, code, and deploy with more control and less friction. New in Kiro CLI 2.0: native Windows support, enhanced subagents, and headless mode for CI/CD automation beyond the terminal.

More Great Builds

Quick hits from the community — demos worth bookmarking:

Allman: Terminal LinkedIn Messenger

Dan Moore • AI Tinkerers - Seattle • Apr 22

Dan Moore built and demoed Allman, a real-time terminal messenger for LinkedIn that uses a Bun-compiled Ink/React TUI backed by a git-versioned JSONL store. The CLI layer handles cookies, rate limits, safety checks, and sending, while the TUI never hit the network directly. Live updates streamed over NDJSON and fs.watch kept the UI responsive, and backfills showed progress counts. People seemed to love the local-first file contract and the clean split between “writer CLI” and “reader UI”. It felt especially timely for agentic workflows where fast context comes from reliable local state, and it could become a practical product for dev teams needing auditable, reversible messaging pipelines.

Loading tech tags...

allman.sh

SAIL: Serverless Agentic Knative Containers

Sébastien Blanc • AI Tinkerers - Paris • Apr 21

Sébastien Blanc, from Sciam, presented SAIL, a serverless, event-driven multi-agent setup on Knative and Kubernetes that wakes agents up on Kafka events and scales them back to zero. In the demo, three agents were deployed via a lean custom resource, then collaborated by sharing context and memory while running locally with Docker Model Runner. It felt especially relevant to today’s push toward agentic workflows that are efficient to operate, and (people really liked how practical it was).

Loading tech tags...

github.comsailtech.orgVideo

Coluur: Claude Cowork App Development

Sabine Pierre • AI Tinkerers - Montreal • Apr 22

Sabine Pierre presented Coluur, a watercolor paint matcher for iOS and web that takes a photo and returns the closest tube matches with pigment codes plus lightfastness and transparency. It uses CIE Lab conversion and a tiny ΔE-based distance function, avoiding model calls for the core matching, then hands off a Notion sprint ticket to Claude Cowork to implement a visible change in the running app quickly. Given her backend and DevOps background at Bell Media, the demo felt like a practical blueprint for solo shipping with guardrails like persistent agent memory. It stood out for combining human-perception color math with agentic execution, and the audience reaction was the kind you can build from.

Loading tech tags...

ruflo: Hive Mind vs Solo Agent

Angie Ariza • AI Tinkerers - Montreal • Apr 22

Angie Ariza, an AI tech lead at Plus Company, presented a live “Hive Mind vs Solo Agent” race where a single Claude Code agent and a ruflo “Tactical Queen” coordinating four specialized workers both built a sliding window rate limiter with tests and 80%+ coverage. The ruflo hierarchy used phased handoffs with Raft to keep work aligned, while the demo streamed progress to a Node.js real-time scoreboard using filesystem polling plus SSE. We liked it because it made multi-agent coordination measurable, and (people loved it) the results pointed toward structured role separation as the practical pattern for faster, higher-quality code. It also hinted at how these orchestrations could productize into reliable coding workflows.

Loading tech tags...

github.com

Eric Chat: Offline Local AI

Eric Fillion • AI Tinkerers - Ottawa • Apr 25

Eric Fillion from Vennify presented Eric Chat, a Python GUI that runs large LLMs locally and offline on Apple Silicon Macs, so users could keep prompts private. Eric Chat paired with the Eric Transformer library to enable quick fine-tuning, pretraining, and RAG workflows through a streamlined MLX-based setup. We liked how it lowered the “getting started” barrier while still feeling technically rigorous, and people (really) leaned in to try it. It also felt aligned with the community push toward efficient local inference that can scale into real tools.

Loading tech tags...

github.comgithub.comVideo

ibaAgent: Agentic Time-Series Analysis

Johannes Garstenauer • AI Tinkerers - Nürnberg • Apr 22

Johannes Garstenauer presented ibaAgent, an agentic time-series analysis CLI built with LangGraph that lets an LLM-driven system inspect and reason over high-resolution signals via tools and scripts rather than raw data. The agent stays mostly blind by using caching, a memory architecture to cope with messy undocumented stores, and a Python sandbox where it dynamically writes analysis code. He tuned prompting and tool use for predictable, interpretable traces. We liked it because it tackles a real bottleneck in long-context agent workflows, and people seemed to enjoy the practical “mostly blind” design as a repeatable pattern for production-bound agents.

Loading tech tags...

github.com

ThreadHop: Tag, Search, and Hand Off Your Coding Agent Sessions

Nanda Kumar • AI Tinkerers - Montreal • Apr 22

Nanda Kumar from Montreal demoed ThreadHop, a local-first TUI and CLI that connects isolated coding-agent sessions into one continuous workspace with chat history, tagging, and cross-session search. ThreadHop indexes local transcripts in SQLite FTS5 and uses compressed handoffs to carry over structured memory like decisions, TODOs, and ADRs into new sessions via a simple skill command. We liked how it kept conversation history from getting lost, especially for builders juggling multiple agent threads, and it felt like a practical follow-through on the broader push toward self-optimizing, hands-off agent workflows.

Loading tech tags...

github.com

NousyBooks: Gemini Live Voice Agents

Vinay Guda • AI Tinkerers - Atlanta • Apr 21

Vinay Guda from Nice Actimize presented NousyBooks, a kids storytelling platform where “Nousy” guides children through bidirectional voice conversations to brainstorm plots, collect character details, and choose art styles. The demo runs a multimodal Gemini Live voice flow, orchestrating 12 function-calling tools to pivot from casual chat into a structured JSON narrative, then uses additional Gemini steps for images and narration, built with React plus a Google Cloud Run and Supabase backend. We liked how he tackled low-latency state and rapid voice interrupts in the genai SDK, and (builders seemed to love it). It fits the current shift toward more autonomous, integrated agents.

Loading tech tags...

github.commedium.com

Endless: Managing Parallel AI Projects

Mike Schinkel • AI Tinkerers - Atlanta • Apr 21

Mike Schinkel presented Endless, a project awareness system that juggles 10+ parallel dev tasks without losing the thread. The demo runs monitors across tmux windows and hooks triggered by prompts and a daemon, then aggregates Claude Code and Codex activity into a local web dashboard using GitHub integration for project state. It was especially compelling because it addressed the real “agent waits too long, so I spin up more” pain, and it kept the next steps visible as dependencies ballooned (people loved it). We liked it for how it mirrors today’s agentic workflows and could grow into a practical ops layer for multi-project engineering.

Loading tech tags...

github.com

Sarathy & me!

Viswa Kumar • AI Tinkerers - Ottawa • Apr 25

Viswa Kumar demoed Sarathy and me, a personal assistant that he built as an OpenClaw-style fork inspired by nanobot, Hermes, and OpenClaw. Sarathy runs as a local-first Python assistant with chat, retrieval, and orchestration that can talk to Ollama and vLLM. He’s the same systems architect behind macOS-native tooling and packet analysis utilities at Dell Canada. We liked it because it made customization feel practical, not just theoretical (people loved the control and setup).

Loading tech tags...

github.com

You are one of 95,000+ readers from Apple, Amazon, Microsoft, Google, Nvidia, OpenAI, Anthropic, Cohere, ElevenLabs, Scale AI, Groq, Mistral AI, and others — spanning frontier labs, big tech, startups, and top universities.