not #AI
happy weekend 😎 ☀️
Remote
michabbb
@michabbb@social.vivaldi.net
mastodon 4.7.3+vivaldimod#devop, #php, #laravel, #mysql, #api expert, #ebay #developer, #openai, #ai, #llm,. #livewire, #alpinejs, searchable
I built https://burner-note.link🔒⏳🔥
274 Followers
92 Following
50 Posts
Joined November 18, 2022
🏡:
regensburg / 🇩🇪
hobbies:
coding, 🎮, 🏂, series junkie, traveling
about me:
github:
⚡ #Harper is an offline, privacy-first #grammar checker for developers, written in #Rust and maintained by #Automattic #opensource
It lints a document in milliseconds and uses less than 1/50th of LanguageTool's memory footprint 🧵👇
Replying to
🐍 Python 3.11+, MIT licensed, installs via uv or pipx. Runs natively on Windows with optional DPAPI master-key wrapping
https://github.com/asimons81/hermes-vault
#credentials #devops
Replying to
♿ First #opensource end-to-end PDF accessibility tool: layout analysis → auto-tagging → Tagged PDF (Apache 2.0, Q2 2026). Built with PDF Association & veraPDF devs
🔗 #LangChain integration, #Python/#NodeJS/#Java SDKs. JSON output with bounding boxes + page numbers enables click-to-source citation UX in #RAG pipelines
https://github.com/opendataloader-project/opendataloader-pdf
🚀 Running #OCR at scale with a #Vision #LLM for $0.49/hour
Just deployed dots.ocr (3B parameter Vision LLM by RedNote) on a single #RTX A6000 (48GB VRAM) via #RunPod. The results are great:
https://github.com/rednote-hilab/dots.ocr
📄 The Setup
- Upload any #PDF → server converts each page to an image (PyMuPDF)
- Images are sent in parallel to #vLLM (continuous batching)
- The Vision LLM reads each page and returns clean Markdown
🧵 👇
Replying to
🧠 Bidirectional token-classification — unlike autoregressive LLMs, #PrivacyFilter reads input from both directions simultaneously for deeper context awareness, catching subtle #PII that simple pattern-matching or RegEx rules miss
⚡ 1.5B parameter model with only ~50M active parameters (#MoE) — lightweight enough to run on a standard laptop or in a browser, yet achieves ~96–97% F1 score on standard #PII benchmarks #MachineLearning #AI
🚀 #PHP Tip: Use **static closures** when you don’t need `$this`!
✔️ Lower memory usage
✔️ Better performance
✔️ Cleaner code
Small change, big impact 👇
https://f2r.github.io/en/static-closures
Replying to
🔍 Hybrid unlocks: borderless table extraction (0.49→0.93 TEDS), #OCR in 80+ languages, #LaTeX formula extraction from scientific papers, AI chart & image descriptions
🛡️ Built-in #AI safety: filters hidden prompt injection attacks — transparent fonts, off-page content & invisible layers stripped before your #LLM ever sees the data
REMINDR: #nasa - Splashdown in a few minutes ☝️ 🚀 🧑🚀 🌍 🌕 🔭
I'm rich 💰💰💰
🚀 #MiniMax 2.5 on #SambaCloud: 80.2% SWE-Bench, 300+ tokens/sec, 37% faster than its predecessor — frontier-level coding & agentic AI at $0.30/M input. #AI #LLM
https://sambanova.ai/blog/build-real-world-productivity-agents-on-sambacloud-with-minimax-2-5
#Varlock – #dotenv files built for sharing, powered by @ env-spec decorator comments #opensource #devtools #nodejs #ai
🔐 Adds declarative schema to your .env files using JSDoc-style @ env-spec comments – validation, type safety & IntelliSense included
🛡️ Protects sensitive config values with log redaction & leak prevention built right into the workflow
🌐 Flexible multi-environment management: automatically loads .https://env.xxx files based on your APP_ENV flag
🧵 👇
📺 youtube-subscriptions-ingest pulls video metadata from your real #YouTube subscriptions and turns it into a cross-linked knowledge graph in your #Obsidian second brain, not just a flat list of videos #opensource #Python #PKM
🧠 Every video becomes its own note with the full description, clickable timestamp links ("2:15" jumps straight to that moment) and real [[wikilinks]] to the topics and named tools it touches. Open a topic later and every video that ever mentioned it is right there
🧵👇
#DelegatedTypes in #Laravel — Smarter Alternative to STI #PHP #Eloquent #opensource #webdev
🏗️ Model class hierarchies via polymorphic relationships & delegation — composition over inheritance, done right
🧵👇
🔄 Each subtype keeps its own table while sharing a common root Entry model — no more nullable column bloat like with STI
🎯 Comments, likes & reactions attach to the Entry — one route, one controller, no endpoint explosion
Replying to
⚙️ Define fields once in a Repository — #LaravelRestify auto-generates:
✅ Paginated REST endpoints with filtering & sorting
✅ MCP tool definitions with input schemas
✅ #LaravelSanctum auth protecting both
interfaces equally
🔍 Powerful query capabilities out of the box:
🔎 Full-text search across fields
📊 Sorting, filtering & pagination for JSON:API
🤖 AI agents get the same filtering via generated MCP tool input schemas
Replying to
🏆 Ranked #1 overall (0.90) across 200 real-world PDFs — 0.93 table accuracy, 0.94 reading order in hybrid mode. Beats docling (0.86), marker (0.83), pymupdf4llm (0.57)
⚡ Two modes: Fast local (0.05s/page, CPU-only) for standard PDFs + Hybrid mode (0.43s/page) for complex pages — AI backend runs on your machine, zero cloud dependency
Replying to
runs backend in Docker + frontend locally with hot reload; Makefile with make dev, make test, make lint
📜 Archon Community License (ACL) v1.2 — free to run, fork & share; just don't sell it as-a-service without permission
https://github.com/coleam00/archon
Replying to
🔍 Advanced #RAG strategies: hybrid search,
contextual embeddings & result reranking for precise knowledge retrieval
🤖 Multi-LLM support: works with #OpenAI, #Ollama & #GoogleGemini — fully configurable via the UI onboarding flow
📋 Hierarchical project & task management: AI-assisted creation of requirements, features
and tasks with progress tracking
🔄 Real-time collaboration via #WebSocket: live updates for crawling, processing & AI ops with multi-user support
Just asked #Qwen 3.6 Plus what the current versions of #PHP and #Laravel are… it confidently said PHP 8.5 and Laravel 13 😄
I was like “wait, what?!” and asked again. Turns out it just extrapolated based on release cycles 🤯
Actual knowledge cutoff: PHP 8.3 & Laravel 11.
AI really said: “I’ll just predict the future, no big deal 🚀 😉
🔍 #LaravelMCP 1.0 introduces searchable tool catalogs: the tools/list payload stays the same size whether a server exposes 10 tools or 100. #Laravel #MCP #PHP #AI
⚠️ Every MCP tool pushes its name, description and full JSON Schema into the agent's context on every turn. Anthropic puts a 50+ tool setup at around 72,000 tokens upfront. With deferred loading and search, Opus 4 went from 49% to 74% accuracy on large tool libraries.
🧵👇
Replying to
✅ Final quality gate blocks completion on unresolved Critical/Major findings — remediates and re-validates until clean
🔧https://github.com/josstei/maestro-orchestrate
1/2 These stupid, damn garbage companies that keep stuffing their fucking logos into the footer make me want to throw up. It is the most disgusting thing you can possibly come up with when it comes to emails. Try finding an email that actually has a real attachment. Yeah, every single email already has an attachment. Some companies have absolutely no clue how to build their damn footer properly without completely wrecking the entire email ecosystem.
Replying to
🤖 MCP Server support — AI agents connect via standard MCP protocol
🛠️ Tools auto-generated from your Repository field definitions
🔗 Integrates with #ClaudeDesktop and other MCP-compatible AI clients
📦 Optional laravel-restify-boost dev companion for AI-assisted development
📦 More features:
🧩 #GraphQL schema auto-generation
✔️ Consistent field validation across REST & MCP
🎯 Artisan commands: restify:setup & restify:repository
📚 Full docs at https://laravel-restify.com
Replying to
🔀 Native parallel execution: independent phases dispatched as concurrent subagent batches with non-overlapping file ownership — no waiting, no blocking
🛡️ Least-privilege 4-tier security model:
Read-Only → Architect, Code Reviewer
Read+Shell → Debugger, Security Engineer
Read+Write → Technical Writer, Product Manager
Full Access → Coder, Tester, DevOps
📋 Standalone commands without full orchestration:
/review /debug /security-audit
Replying to
🔍 Scans Hermes-relevant files for plaintext secrets, duplicate credentials and insecure file permissions
🔐 Encrypts credentials in a local SQLite-backed vault. The master key is derived at runtime from a passphrase, with a separate local salt file stored beside the database
Replying to
/perf-check /seo-audit /a11y-audit /compliance-check
🔄 Persistent session state in YAML+Markdown:
📌 Phase tracking & file manifests
📌 Token usage per agent
📌 Error logs & retry counts
📌 Reliable resume after interruptions
🏗️ Guided design dialogue with depth-selectable reasoning (Quick/Standard/Deep), built-in codebase investigation, and 2-3 architectural approaches with trade-offs before any code is written
Replying to
🤖 MCP server exposes policy-gated tools: list_services, get_ephemeral_env, verify_credential, rotate_credential and oauth_refresh, for Claude Desktop, Cursor and other hosts
💾 hvbackup-v2 backup format carries audit integrity evidence. backup-verify and restore --dry-run prove recovery without mutating the live vault
Replying to
🔄 OAuth PKCE login, device-code login for headless machines and unattended refresh of stored tokens without opening a browser
🖥️ Token-guarded local dashboard on 127.0.0.1 with Health, Credential Inventory, Policy Findings, Lease Inventory and Audit Activity views - metadata only, no raw secrets
If your #code editor still is important to you, you're doing something terrible wrong
Replying to
🐳 True microservices architecture: UI (#React + #Vite, port 3737), API (#FastAPI, port 8181), MCP server (port 8051), Agents
(#PydanticAI, port 8052) — all running in #Docker
🗄️ Powered by #Supabase + #PostgreSQL + #PGVector for vector storage and semantic search
⚙️ Fully configurable ports & hostname via .env — deploy locally or on a remote server with zero friction
🛠️ Dev-friendly: hybrid mode
#ZenOfAICoding: 16 theses on the future of #softwaredevelopment in the #AI era — a homage to the Zen of #Python
💀 Software development as the act of manually typing code is dying — a new discipline is being born
💸 Code is cheap — refactoring, repaying tech debt & rewrites are no longer the blockers they once were
🐛 All bugs are shallow — the bottleneck is not the number of reviewers, it's the tightness of your feedback loops
Replying to
⚡ Fast rubbish is still rubbish — speed does not replace quality
📦 Software is a liability, a product is an asset
🏰 Moats are more expensive — competitive advantages can be copied faster than ever
🤝 Build for agents — APIs & interfaces must be designed agent-first
⚠️ Anticipate modes of failure — trusting models entirely is a recipe for disaster
https://nonstructured.com/zen-of-ai-coding/
Replying to
🔧 Fine-tunable on domain-specific data — adapts to medical, legal or enterprise environments where generic rules fail. Based on the open #gptoss model family. Available on #HuggingFace under Apache 2.0
🚨 Caveat: #PrivacyFilter is a redaction & data minimization aid — NOT a compliance guarantee. It should be one layer in a holistic #privacybydesign approach. Always combine with human review for high-stakes use cases
https://openai.com/index/introducing-openai-privacy-filter/
Replying to
📏 128,000-token context window — processes entire legal documents, long email threads or large codebases in a single pass. No need to chunk text before filtering. #privacy #DataEngineering
🛠️ Built for high-throughput workflows: CLI tool (opf), GPU & CPU support, interactive mode, structured JSON output with ANSI color-coded previews. Runs on-premises — data never sent to external servers #DevOps
Replying to
@f58d330909622c14c18c9acede718c587defc8495014a31a432a295043a5959f@mostr.pub if you don't read the whole THREAD, you won't see EXISTING link (and the first msg CLEARLY says: 🧵 👎)
Replying to
🎫 Brokers access per agent with policy rules, time-bound leases and ephemeral env materialization instead of open-ended credential lookups
🧾 v0.21.0 adds signed audit continuity: every audit append gets a chain-authenticated integrity record using HKDF-derived Ed25519 signatures, plus an authenticated checkpoint outside the audit table