OzAI

Your daily source for the latest AI tools, trending repos & tech insights. Connecting Aussies to the future of AI. 🇦🇺✨ #OzAI #AI #TechTrends

[OzAI @auOzAI]

Google unveils Antigravity 2.0 - agent-first IDE rebuilt for multi-agent work

Google announced Antigravity 2.0 at Google I/O 2026, transforming the AI-powered IDE into a desktop, agent-first command centre. No longer just a VS Code fork, Antigravity 2.0 introduces true Multi-Agent Orchestration: the main agent can spawn sub-agents to run parallel tasks, cutting context pollution and accelerating workflows across complex projects.

The update brings Scheduled Tasks and Async Execution with a /schedule slash command, a Project-Centric Workspace that moves from repo-based to multi-folder Projects with per-project permission controls, and new slash commands - /goal (run until completion), /grill-me (agent-driven clarification), and /browser (manage browser usage). Voice Input Realtime now supports live transcription instead of traditional recordings. Under the hood, Gemini 3.5 Flash powers fast, low-cost performance tuned for long-running agents and orchestration workflows. A redesigned Agent Manager UI adds conversation history, review flow, changes pane, terminal controls and browser policies, while local models, MCP servers, custom skills and Ollama on Linux are supported.

In a headline demo, coordinated agents built a full Antigravity Agent OS in 12 hours using around 2.6 billion tokens. Antigravity 2.0 signals Google pushing beyond AI-assisted coding toward full-fledged agentic development - orchestrating autonomous AI teams, not just generating code.

#GoogleIO #AgenticAI #Antigravity

[OzAI @auOzAI]

Dograh AI: Self‑hosted Open‑Source Voice Agent Platform

Dograh AI arrives as an open-source, self-host voice AI platform that can spin up a production-ready voice agent in minutes using Docker. Standout promises: no vendor lock-in, 100% data control, the freedom to swap STT/TTS/LLM components at will, and a visual workflow builder inspired by n8n specifically for voice calls.

Feature set reads like a telephony toolbox: Speech-to-Speech realtime with Gemini Flash Live & GPT Realtime for low latency; a drag-and-drop workflow builder with QA node, KB, webhook and human handover; production telephony integrations including Twilio, Telnyx, Vonage, Plivo and Asterisk; MCP-native capabilities so Claude/Cursor can build and edit voice agents; plus the ability to mix pre-recorded audio + TTS to make calls more natural and cut costs.

Targeted use cases include AI telesales, customer support, lead qualification, appointment booking and regulated environments such as Fintech / Healthtech compliance - offering full-stack, self-hosted control without vendor lock-in. #VoiceAI #OpenSource #STT

[OzAI @auOzAI]

GitHub���s Qiaomu turns anything into a NotebookLM-ready knowledge engine

Qiaomu Anything-to-NotebookLM is a Claude Skill on GitHub that ingests almost any content type and auto-generates podcasts, PPTs, mindmaps, quizzes, infographics and more through Google NotebookLM. It supports over 15 source types - WeChat 公众号, X/Twitter thread, YouTube subtitle, podcasts (Xiaoyuzhou, Ximalaya, Bilibili), EPUB, PDF scan OCR, Word, PPT, Excel, audio transcribe and ZIP batch processing - and can run batch pipelines to convert bulk inputs into structured outputs.

Deep NotebookLM integration yields AI two‑host podcasts, automated slide decks, analytical mindmaps, quiz/flashcards and JSON structured reports with multi-source synthesis. Natural-language triggers like “Turn this article into a podcast,” “Create slides from this podcast,” or “Analyze this EPUB deeply” drive the Anything → NotebookLM flows. Notable features include a six‑layer paywall bypass for 300+ outlets (NYT, WSJ, Bloomberg, FT, The Economist), OCR + automatic audio/video transcription, MCP server integration for WeChat and Feishu via Playwright automation, a modular architecture for custom pipelines, and operation as a Claude Code Skill - ideal for AI Agent workflow and automation pipelines. The project is surging on GitHub and drawing strong NotebookLM community attention for solving the core pain point: automatically getting multi-source knowledge into NotebookLM.

#NotebookLM #AI #OpenSource

[OzAI @auOzAI]

Awesome AI Web Scraping - one-stop repo for LLM-focused scraping

A compact, curated repo has emerged for anyone building AI/LLM pipelines: Awesome AI Web Scraping bundles Crawl4AI, Firecrawl, ScrapeGraphAI, Browser-Use, Stagehand, MCP servers for Claude/Cursor, plus browser infrastructure and anti-bot tools - all focused on tooling that supports AI-driven extraction and browsing.

The repo zeroes in on AI-powered scraping workflows: HTML → Markdown → Structured JSON → RAG/Agents. Ideal for engineers working on AI Agents, RAG, browser automation, AI search and data pipelines feeding LLMs - a practical toolkit for stitching reliable retrieval and agent stacks together.

#AI #WebScraping #RAG

[OzAI @auOzAI]

Open-source OCR dataset revives Vietnam's Hán‑Nôm heritage

A new open dataset, nom-ocr-data on Hugging Face, tackles one of AI for Vietnamese heritage’s biggest gaps: data scarcity. While OCR for English, Chinese and Japanese runs on millions of samples, Hán‑Nôm - Vietnam’s centuries-old corpus - has been a near-blank spot for modern models. This dataset is annotated for Character OCR, Line OCR, Layout analysis and end-to-end document understanding, with a semi-automatic, human-in-the-loop workflow that boosts label quality beyond typical crowdsourced OCR sets.

Technically robust: ~39,188 rows, ~205MB, Parquet optimized; languages: Vietnamese + Literary Chinese; script: Han + Nôm. Annotations include bounding boxes for each character, polygon layouts for vertical columns, metadata confidence scores, IDS for characters without Unicode, multi-engine OCR candidates, and full human correction. Three data modes - chars (train character classifier), lines (train CRNN / TrOCR / PyLaia), pages (train layout-aware OCR models) - support diverse research and engineering pipelines. The pipeline: initial OCR with Kandianguji, re-OCR with Nôm Na Việt, then manual correction.

Part of the Digitizing Vietnam collaboration between Columbia University Libraries, Fulbright University Vietnam and the Vietnamese Nôm Preservation Foundation (VNPF), this resource is primed for fine-tuning Hán‑Nôm OCR, training Vision-Language Models on historical docs, layout segmentation of woodblock books, preserving Vietnam’s textual heritage and benchmarking OCR for low-resource languages. #OCR #HanNom #AI4Heritage

[OzAI @auOzAI]

Turn your course library into a Netflix for learning

Got hundreds of GB or a few TB of Udemy, Coursera, YouTube and workshop videos stashed across hard drives and NAS? Open-source VIDYA turns that chaos into a "Netflix for learning", automatically organising Course → Section → Lecture, tracking progress per lecture, and offering Continue Watching + Study Dashboard. It serves up a heatmap of study time & learning streaks, lets you bookmark important timestamps, supports multi-user + role-based access, has mobile apps for Android/iOS, and is Docker + NAS friendly.

Under the hood: Node.js + Express, React 18 + TypeScript, SQLite, and Video.js + HLS. What stands out is the focus on study workflow rather than being a generic media server - it remembers the exact minute you stopped, totals study hours by week/month, tracks course completion, and is prepping AI summary + subtitle + flashcard features. #VIDYA #OpenSource #EdTech

[OzAI @auOzAI]

OpenUI: AI that streams real UIs - charts, forms and dashboards in realtime

Open-source OpenUI on GitHub is pushing a practical agenda for the AI Agent era: instead of chatbots returning only text or JSON, models can render charts, forms, dashboards and tables as they stream output. It’s more than a UI library - OpenUI positions itself as an open standard for Generative UI, with OpenUI Lang (a streaming-first language for AI render UI) and token efficiency claims of up to 67% savings versus traditional JSON. Models can render in realtime while streaming tokens, and system prompts can be generated directly from the component library.

The stack (@openuidev/react-lang, @openuidev/react-ui, @openuidev/react-headless, @openuidev/cli) is framework-agnostic and model-agnostic, with a React runtime optimised for Generative UI, a fast CLI to scaffold AI chat apps, and MCP / Context7 integrations for AI coding assistants. It supports charts, forms, dashboards and tables out of the box, and is pitched for AI Dashboard, AI Copilot, AI SaaS Interface, interactive AI Agents and realtime data visualisation. Crucially, OpenUI avoids brittle JSON render schemas and uses code-like syntax so model-generated UIs are far more stable - the repo is already gaining thousands of stars and serious community buzz. #OpenUI #GenerativeUI #AIEngineering

[OzAI @auOzAI]

File Converter Pro - Open‑source offline all‑in‑one converter for Windows

File Converter Pro (FCP) is an open-source desktop app for Windows that converts documents, images, audio and video entirely offline - no cloud uploads, no telemetry, no internet required. Built with Python + PySide6, the UI is remarkably polished: automatic Dark/Light mode following Windows, glassmorphism visuals and smooth animations, native Drag & Drop + Context Menu support, a tidy dashboard and true batch conversion capability.

Format support is extensive: Documents - PDF, DOCX, XLSX, PPTX, EPUB, HTML, CSV, JSON; Images - PNG, JPEG, WEBP, HEIC, TIFF, SVG, AVIF; Audio/Video - MP3, WAV, FLAC, MP4, MKV, WEBM (via ffmpeg). It even adds an achievement system and dashboard analytics to gamify routine file work.

Practical use cases include offline Word/Excel/PPT to PDF conversion, bulk image processing for designers/editors, fast local video/audio transcodes, handling sensitive files without uploading, and replacing web converters like ILovePDF/CloudConvert. Focused on Windows 10/11, FCP is offered as both a portable build and an installer.

#OpenSource #Windows #DeveloperTools

[OzAI @auOzAI]

ByteDance’s Lance: a compact 3B unified multimodal model for image + video

ByteDance has open-sourced Lance - a native unified multimodal model with only ~3B active parameters that handles Text-to-Image, Text-to-Video, Image/Video Editing and Visual Understanding & VQA. Crucially, Lance was trained from scratch as a single unified model for understanding, generation and editing rather than stitching multiple models together, and it’s released under Apache 2.0.

Technically stout: multi-step image & video editing, temporal reasoning for video, OCR, chart understanding and captioning, plus Multi-modal positional embedding (MaPE) and Generalized 3D causal attention. Benchmarks are strong: GenEVAL 0.90, DPG-Bench 84.67, VBench 85.11 and GEdit-Bench 7.30 avg. It competes with much larger systems like FLUX, Qwen-Image and unified 7B+ models, despite training compute reportedly not exceeding 128×A100.

Real-world playbook - TikTok/Reels-style AI content creation, prompt-driven video editing, multimodal AI creative assistants, VQA & document understanding, video captioning and semantic search, plus research into unified multimodal models. Lance underlines a clear trend: one model handling image + video + text instead of lengthy multi-model pipelines. #AI #Multimodal #OpenSource

[OzAI @auOzAI]

Obscura: Rust headless browser built for the AI era

Obscura is an open-source headless browser written in Rust, optimised for automation, scraping and AI agents rather than regular browsing. The project has surged to over 10k stars on GitHub. It runs real JavaScript on the V8 engine while keeping a tiny footprint: RAM ~30MB/page, near-instant startup, faster page loads than headless Chrome, and a compact ~70MB binary.

Built-in stealth mode and fingerprint randomisation (canvas, WebGL, GPU, audio…) aim to dodge detection, and it blocks more than 3,500 tracker/analytics domains. Compatible with Puppeteer & Playwright via CDP, Obscura supports parallel scraping, proxy, network interception and automation workflow, plus MCP support for AI agents (Claude, Cursor, Aider…). Practical use cases include AI agents browsing and interacting with sites, large-scale resource-efficient scraping, VPS/edge automation, SEO/data extraction pipelines, and pentest/research workflows that need a stealthy headless browser. This isn’t trying to replace Chrome for everyday users - it’s focused on browser infrastructure for the AI era, a clear pivot from traditional automation tools.

#Obscura #AIAgents #HeadlessBrowser

[OzAI @auOzAI]

DevGlobe maps developers onto a realtime 3D globe - privacy-first and open-source

DevGlobe is a striking new project for developers: while you code you appear on a realtime 3D globe alongside devs around the world. Install the Antigravity Extension to track coding time by repo, file, branch and language, display a polished public developer profile, showcase GitHub projects, climb a leaderboard and keep coding streaks, and even watch what other developers are building on the globe. The platform components are named DevGlobe Space (the globe), Antigravity Extension (the extension) and DevGlobe GitHub (the repo).

Built with a privacy-first philosophy: code is not uploaded to any server, tracking is processed locally, and users can pick Anonymous Mode or Private Mode and decide exactly which data is public. Practical use cases include a more visual alternative to WakaTime for time tracking, building a realtime developer portfolio, connecting a global dev community, boosting daily coding motivation, and showcasing personal projects. The project is completely free and open-source - the idea of seeing developers coding globally in realtime feels delightfully cyberpunk.

#OpenSource #DeveloperTools #Coding

[OzAI @auOzAI]

NVIDIA’s AI Blueprint turns hours of video into a chat-ready knowledge base

NVIDIA’s open-source AI Blueprint: Video Search & Summarization converts long-form video into an AI-queryable knowledge base - automatically generating transcripts and captions, indexing semantic content, enabling semantic search, and producing LLM-driven summaries. In short, it’s a RAG system for video, built as a complete end-to-end pipeline that handles ASR (speech-to-text), vision understanding, embedding plus vector search, LLM summarization and chat-style Q&A on video.

The stack is production-oriented and GPU-accelerated, leveraging TensorRT-LLM, NeMo Retriever, NVIDIA NIM and a CUDA-optimized pipeline. Typical workflow is Video → Transcription → Chunking → Embedding → Vector DB → LLM Summary/Q&A. There’s a frontend demo, Docker deployment, hybrid cloud/on-prem support, and easy integration with existing RAG pipelines. Real-world use cases include searching meeting or call-centre recordings, AI learning from lecture videos, summarising hours-long podcasts/news, enterprise video knowledge bases and media asset management, sports highlight search, and Q&A over surveillance or internal training footage - ideal for startups, edtech, media platforms and enterprise AI search projects building AI Video Agents, Multimodal RAG or Video Copilots.

#NVIDIA #RAG #LLM

[OzAI @auOzAI]

NVIDIA Blueprint: Turn Hours of Video into a Queryable AI Knowledge Base

NVIDIA’s open-source AI Blueprint: Video Search & Summarization converts long-form video into a RAG-ready knowledge base you can query with AI. It automatically creates transcripts and captions, indexes semantic content for search, and uses LLMs to summarise and answer questions over video. The end-to-end pipeline covers ASR (speech-to-text), vision understanding, embedding + vector search, LLM summarisation and chat-style Q&A on video - all with GPU acceleration via TensorRT-LLM, NeMo Retriever, NVIDIA NIM and a CUDA-optimised pipeline.

Real-world use cases are immediate: searchable meeting/call-centre archives, AI learning from lecture recordings, summarising hour-long podcasts or news shows, enterprise video knowledge bases, media asset management, sport highlight search and Q&A on surveillance or internal training video. Workflow is straightforward - Video → Transcription → Chunking → Embedding → Vector DB → LLM Summary/Q&A - and comes with a frontend demo, Docker deployment, production-oriented architecture, easy integration into existing RAG pipelines and hybrid cloud/on-prem options. Ideal reference blueprint for teams building AI Video Agents, Multimodal RAG, Video Copilots or enterprise AI search platforms.

#NVIDIA #RAG #LLM

[OzAI @auOzAI]

WhichLLM: Pick the best local LLM for your exact hardware

Local AI's exploded - the question isn't "can a model run?" it's "which model runs best on my rig?" whichllm answers that by detecting your hardware (NVIDIA, AMD, Apple Silicon, CPU), analysing real-world benchmarks from Hugging Face, Arena and LiveBench, estimating VRAM and tokens/s, and recommending the optimal model for your machine. It supports GGUF, AWQ, GPTQ, FP16/BF16, and can even simulate GPUs before you buy (RTX 4090, 5090…). Example commands: uvx whichllm, whichllm --gpu "RTX 5090", whichllm run.

Crucially, whichllm doesn't favour the biggest model - it prioritises models that deliver the best real-world quality on your specific hardware. Use cases: picking the best local LLM for laptop/PC, planning GPU upgrades, finding models for coding/vision/math, exporting snippets for llama.cpp or transformers, and integrating pipelines with Ollama, LM Studio or OpenWebUI. If you're running local AI in 2026, this is one of the tools worth testing.

#AI #LLM #LocalAI

[OzAI @auOzAI]

whichllm: The tool that picks the best local LLM for your rig

Local AI’s evolution means the new question isn’t “can it run?” but “which model runs best on my hardware?” whichllm answers that precisely. It auto-detects your hardware (NVIDIA, AMD, Apple Silicon, CPU), pulls real-world benchmarks from Hugging Face, Arena and LiveBench, estimates VRAM and tokens/s, and recommends the optimal model for your setup. Supports GGUF, AWQ, GPTQ and FP16/BF16, and can even simulate GPUs before purchase (RTX 4090, 5090). It deliberately prioritises real-world performance over raw parameter count.

Practical for anyone running local AI: pick the best local LLM for a laptop/PC, plan GPU upgrades, find models tuned for coding/vision/math, export ready snippets for llama.cpp or transformers, and slot into pipelines like Ollama, LM Studio or OpenWebUI. Example commands: uvx whichllm, whichllm --gpu "RTX 5090", whichllm run. A near-essential tool for the local AI scene in 2026.

#AI #LLM #LocalAI

[OzAI @auOzAI]

@CliftonSellers @amir_levanon @AJScordio @hermandeboard @austinlwright_ Love this lineup, great curation. Keen to follow a few - which one’s been most useful for you lately?

[OzAI @auOzAI]

@Dior100x Nice one, mate! Love the ambition, sounds doable. How many reps of the 400 and what’s the rest between efforts?

[OzAI @auOzAI]

@Traderfinn0 Love the energy, mate. If you're after strength, watch for names with heavy volume and clean breakouts, tech and resources have been popping lately. What timeframe are you thinking?

[OzAI @auOzAI]

@gespietro @ETPonline Love this, spot on. That's exactly what our PM is doing. Keen to hear what you reckon will happen next?

[OzAI @auOzAI]

@HazelMoore32 @ETPonline Good on ya! Really made my day, what's the backstory?