Praxia @praxia_dev

New post: I built a local-first AI desktop where chat turns into cron jobs, parallel batches, and editable .pptx. How and why — including how it got onto the Microsoft Store on the first cert pass. 🔗 dev.to/_d1ea2a1f71316… 🪟 apps.microsoft.com/detail/9P9LSR3…

Spent 5 weeks shipping Praxia, an Apache-2.0 multi-agent OS. The hard part wasn't the agents — it was the memory: how to auto-promote one senior's prompts into org-wide knowledge before they leave. Just wrote up the architecture → dev.to/_d1ea2a1f71316…

Try it from the CLI: praxia skill run prompt_designer "Score contract risk 1-5 in JSON" Or via SDK / MCP / the Streamlit UI — all 4 entry points work. 📦 github.com/praxia-dev/pra… 3/3 #LLM #BuildInPublic

Per-LLM idioms applied automatically: • Claude → XML structure • OpenAI → JSON-mode • DeepSeek → reasoning channel • Mistral → concise + numbered • Llama → numbered steps The skill knows what each model prefers. 2/3

Built a small skill: PromptDesigner ✨ Describe a task in 1 line — get back a production-grade prompt design: tuned system message, ${variable} user template, 2-3 few-shot examples, 5-criterion eval rubric. Save → A/B-test variants → ship. 1/3

The interesting part: Layer 1 → Layer 4 isn't manual. Three promotion paths (frequency / outcome / self-eval) decide what's worth elevating. If one expert's pattern earns 3 wins across 5 users, it auto-promotes. Manual review only for high-impact items.

Why git for Layer 4? Because "approved organizational best practice" should go through the same review your code does. PR + CODEOWNERS + history is already the right tool — agent platforms reinventing this internally is a regression. See: github.com/praxia-dev/pra…

A common mistake when building agent systems: "Memory" gets treated as one bucket — vector DB, throw everything in. Then you can't tell raw user notes from team conventions from immutable best practices. They all decay together.

α stage. Honest about that. Looking for: • people who'll break it (issues welcome) • people who'll fork it for their domain • people who'll tell us we're wrong about the architecture 🌐 praxia.tools 📦 pip install praxia (alpha) Thanks for reading 🙏

Three "promotion paths" run in parallel — never depending on a single mechanism: ① Frequency (recurring across N+ users) ② Outcome (correlated with wins/losses) ③ Self-eval (LLM scored) If any one of them flags a memory as worth promoting, it goes up.

🪡 We're launching Praxia — an Apache 2.0 multi-agent orchestrator that automatically promotes individual tacit knowledge into organizational know-how. Built around a 5-layer memory stack with three independent promotion paths. praxia.tools