Community Spotlights • Issue #12

Community Spotlights • Issue #12

Issue #12 · Week of December 29

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 that stood apart from the 70+ demos presented at our meetups around the world over the last two weeks, and found clear patterns of builders focusing heavily on reliable agentic workflows and sovereign, local-first AI architectures. Benjamin Shoemaker (Seattle) showed Vibe Scaffold, generating structured specs for reliable coding agents. Dhar Rawal (Houston) demonstrated fastworkflow beating SOTA on Tau Bench using small models. Alula Zeryihun (Houston) unveiled Zulu.cash, a zero-cloud blueprint for private, local-first agents. Congrats to these builders!

tbai: A mini HuggingFace for robots Jakub Jon Student at FEE CTU 📍 AI Tinkerers - Prague • Dec 16 Jakub Jon, a student at CTU FEE, presented tbai, a mini HuggingFace for robots. tbai provides a full-stack robotics pipeline from inter-process communication and state estimation to cluster-scale policy training, with validation in simulation and real-world deployment. Its scope and tooling (ROS, Gazebo, C++, Python) let researchers build and deploy legged-robot controllers quickly, and early survey feedback noted strong practical appeal. This is a glimpse of production-ready robotics infrastructure with potential to scale for labs and startups prototyping embodied AI. Read more → TECH STACK PyTorchHuggingFaceONNXPythonC++ PROJECT LINKS github.com github.com

NVIDIA Cosmos: Lab to Field Paula Ramos Computer Vision and ML Advocate Lead at Voxel51 📍 AI Tinkerers - Raleigh • Dec 10 Paula Ramos, Senior DevRel at Voxel51, presented From Lab Walls to Real Fields: Fixing Data Scarcity with NVIDIA Cosmos-Transfer2.5 + FiftyOne. The demo shows BioCLIP selecting rare BioTrove moth samples and turning them into controllable video inputs, while Cosmos-Transfer2.5 translates lab images to field-like scenes and FiftyOne validates results. It yields 20-40% improvements on downstream classifiers by expanding rare classes and correcting bias, and the pipeline is fully reproducible and open-source. Survey feedback noted its practical value for real-world datasets. Read more → TECH STACK Cosmos-Transfer2.5Real2RealFiftyOneBioCLIPBioTrove Dataset PROJECT LINKS voxel51.com nvidia-cosmos.github.io Video

Train Guard Nikita Chistyakov Payment/Data Analyst at Back Market 📍 AI Tinkerers - Paris • Dec 09 Nikita Chistyakov presented Train Guard, a project that preprocesses accident reports (PDFs) to build a graph database linking causes and consequences. It ingests PDFs, uses a Python-based NLP pipeline to extract entities, and constructs a knowledge graph where edges carry weights that reflect relationship strength. The approach is visually intuitive, and the audience noted its clear grounding for complex relationships. For builders, it shows how domain knowledge can be encoded into a compact graph for scalable reasoning. Read more → TECH STACK Neo4jLLM Graph Builder MCPGPT-4LangChainOpenAI API PROJECT LINKS

AI: Organizational Context Translation Stephen F. Johnston Principal at Gambasta, LLC 📍 AI Tinkerers - Atlanta • Dec 16 Cooking Read more → TECH STACK Loading tech tags... PROJECT LINKS github.com google.com @glstephen

Aysnc AI Coding Lee James Senior SWE at Google 📍 AI Tinkerers - Atlanta • Dec 16 Lee James, a senior SWE at Google DeepMind, presented Async AI Coding, a demo of running multiple coding agents concurrently with Claude Code via GitHub Actions and a CI/CD pipeline that coordinates their work. The implementation emphasizes asynchronous orchestration and scalable task execution, with concrete examples of agent turns and artifact reviews. It’s a practical peek at production-ready multi-agent coding workflows, hinting at teams harnessing automated code generation at scale. It’s a practical blueprint for parallel AI tooling in practice. Read more → TECH STACK Claude CodeGitHub ActionsPythonTypeScriptCI/CD PROJECT LINKS

Read more »

Sponsored by Vanta

🌍 The startup guide to global compliance

Growing? This guide helps you cut through SOC 2, ISO 27001, GDPR, and more so you only focus on what matters.

More Great Builds

Quick hits from the community — demos worth bookmarking:

Minimal Policy-as-Code Rule Engine

Athena Biju • AI Tinkerers - Bremen • Dec 10

Athena Biju from Jacobs University presented Encoding Regulation, a policy-as-code engine for transaction screening. She walked through loading YAML rules, parsing them into an AST, and evaluating transactions via a pipeline to yield allow, block, or flag decisions. The demo included a sandbox for rule tweaks and a setup with an HTML interface. Audiences noted its evaluation semantics and traceable decisions, with potential to pair symbolic rules with learned models for finance, a reminder that rule engines are within reach.

Loading tech tags...

github.comaspasia-pilot.streamlit.app

FlowCoder: Visual Agent Workflows

Vighnesh Leonardo Shiv • AI Tinkerers - Seattle • Dec 08

Vighnesh Leonardo Shiv from PX, a Caltech-trained AI researcher, presented FlowCoder, a visual flowchart builder for configuring autonomous workflows with Claude Code and Codex. FlowCoder uses directed graphs with branches and loops, a variable-scoped command stack, and built-in Git snapshots after each block, all orchestrated by Python/Bash. It builds on top of existing coding agents rather than raw LLM APIs. (peers liked the visual approach) and the setup hints at production-ready workflows for long-horizon automation.

Loading tech tags...

github.com

Vibe Scaffold - A wizard-style tool that guides you from a rough idea to a refined technical architecture that AI coding agents can execute

Benjamin Shoemaker • AI Tinkerers - Seattle • Dec 08

Benjamin Shoemaker from Seattle presented Vibescaffold, a wizard-style tool guiding ideas to refined architecture for AI coding agents. The demo shows vibescaffold.dev, which generates structured DEV_SPEC and PROMPT_PLAN documents to anchor agent context, and a Next.js wizard using gpt-5-nano to create context-aware agent specs. Audience feedback indicated it resonated for practical, reliable agent tooling (people loved the pragmatic approach). It points to a business potential: a production-ready planning framework that speeds up AI coding workflows.

Loading tech tags...

vibescaffold.devgithub.com

AgenticInternet

Chrys NIONGOLO • AI Tinkerers - Paris • Dec 09

Chrys NIONGOLO presented AgenticInternet, which automates complex online workflows by building AI agents that autonomously browse, interact, and collaborate. The demo showcases Python-based agents using specialized search APIs to execute long-horizon tasks, with a lean orchestration approach. Survey feedback noted its solid execution and clear problem framing, hinting at practical business value by reducing manual toil. If productized, it could empower teams to run end-to-end tasks with minimal human input. Builders can take away the power of scalable agent orchestration.

Loading tech tags...

github.commermaidchart.com

ChatGPT: Building Agentic Apps

Quentin Churet • AI Tinkerers - Paris • Dec 09

Quentin Churet from Alpic presented Building agentic experience in ChatGPT Apps: MCP Meets UI, showing how to adapt an MCP Server to expose a UI and interact with the agent from your app. He demonstrated Skybridge, a TypeScript framework for building typed React widgets for ChatGPT tool execution, and how MCP Apps render interactive interfaces inside workflows. The session drew positive feedback for practical code and opened a path to production-grade agent UIs for wider use.

Loading tech tags...

github.comalpic.ai

fastworkflow: Small Models, SOTA Agents

Dhar Rawal • AI Tinkerers - Houston • Dec 09

Dhar Rawal from Radiant Logic presented fastworkflow, showing it reached SOTA Tau Bench results with 7B–14B models and even beat Sonnet 4.5 on retail and airline benchmarks. fastworkflow builds deterministic, fault-tolerant, conversational AI workflows using Python, DSPy, and LLMs. It is open-source here: https://github.com/radiantlogicinc/fastworkflow, and feedback from viewers suggested strong interest (people loved it) as it enables reliable agentic work at small scale. For builders, it shows how production-ready, cost-efficient agentic workflows can scale with modest models.

Loading tech tags...

github.comVideo

Zulu.cash: Private Local Agent

Alula Zeryihun • AI Tinkerers - Houston • Dec 09

Alula Zeryihun from Edge Consulting Labs showcased Zulu.cash: Local-First Private AI Agent (ZEC-Native), an on-device agent that handles live transcription, reasoning, and memory storage without sending raw data to the cloud. It runs on Ollama with SQLCipher and WhisperX for private memory, uses MPC and ZEC-native identity, and triggers actions on-device via MCP with minimal slides. This contrasts SaaS-backed agents and hints at sovereign, privacy-preserving AI for developers. For builders, it demonstrates a producible on-device agent OS pattern.

Loading tech tags...

github.comzulu.cashVideoArXiv Paper

Dlab-852-Mini Cultural Alignment

Devansh Gandhi • AI Tinkerers - Hong Kong • Dec 18

Devansh Gandhi, founder and CTO of Decisions Lab, presented Culturally Aligned AI: Building Dlab-852-Mini for Hong Kong Cultural Nuances. The talk showcased a fine-tuned Phi-3.5-mini-instruct model aligned to Hong Kong culture, driven by CultureKit eval and a CLI to measure cultural biases. It highlighted benchmarks with up to 2x better local response fidelity vs the base model, with open-source footprints in culturekit and HuggingFace Dlab-852-Mini. A practical, ethically localized AI direction for global teams.

Loading tech tags...

github.comhuggingface.co

You are one of 95,000+ readers from OpenAI, Anthropic, Google DeepMind, Mistral AI, Cohere, ElevenLabs, Perplexity, Midjourney, Scale AI, Together AI, Groq, Weaviate, and others — spanning frontier labs, big tech, startups, and top universities.