Hacker News: nonatofabio

New comment by nonatofabio in "Microsoft-Decision-1, our model for fast decision-making"

nonatofabio — Fri, 09 Oct 2026 23:40:02 +0000

Yep, Fabio here, and I agree!

Show HN: Luna Agent – Custom AI agent in ~2300 lines of Python, no frameworks

nonatofabio — Sat, 07 Mar 2026 20:49:53 +0000

Hey HN. I evaluated three agent frameworks for a homelab project, one had 400K lines of code (and 42K exposed instances on Shodan), one was 9 days old, and the third was so thin I'd rebuild most of it anyway. None worked for me.

So I built my own in ~2300 lines of Python. No frameworks, 8 runtime dependencies, 106 tests.

What it does: - Persistent memory via SQLite (FTS5 keyword search + sqlite-vec embeddings + recency decay, fused with Reciprocal Rank Fusion) - MCP tool integration — add capabilities by editing a JSON file - Native tools with safety guardrails (bash blocklist, timeouts, output caps) - Discord interface with session isolation - Structured JSON logging for every operation - Conversation compression for effectively infinite context

Runs locally on 2x RTX 3090 with Qwen3-Coder-Next via llama-server. No cloud APIs.

The design philosophy was: don't build what you don't need, but don't block the insertion points. For example, my AI firewall isn't built yet, but all LLM traffic goes through a single configurable URL, swapping in a filtering proxy is a config change I'll do later.

DESIGN.md documents the reasoning behind every architectural decision. Tests mock the LLM client so you can run them on a laptop.

GitHub: https://github.com/nonatofabio/luna-agent Blog post with full technical deep-dive: https://nonatofabio.github.io/blog/post.html?slug=luna_agent

Happy to answer questions about any of the design tradeoffs.

Comments URL: https://news.ycombinator.com/item?id=47291343

Points: 1

# Comments: 0

Show HN: Local_faiss_MCP – A tiny MCP server for local RAG (FAISS and MiniLM)

nonatofabio — Wed, 03 Dec 2025 16:22:11 +0000

I built this because I got frustrated with the current state of "local" RAG. It felt like I had to spin up a Docker container, configure a vector DB, and manage an ingestion pipeline just to let Claude ask questions about a few PDFs in a folder.

We seem to have turned "grep with semantics" into a microservices architecture problem.

What this is: local_faiss_mcp is a minimal implementation of the Model Context Protocol (MCP) that wraps FAISS and sentence-transformers. It runs entirely locally (no API keys, no external services) and connects to Claude Desktop via stdio.

How it works:

You run server.py (Claude runs this automatically via config).

It uses all-MiniLM-L6-v2 (on CPU) to embed text.

It stores the vectors in a flat FAISS index on disk alongside a JSON metadata file.

It exposes two tools to the LLM: ingest_document and query_rag_store.

The stack:

Python

mcp (Python SDK)

faiss-cpu

sentence-transformers

It’s intended for personal workflows (notes, logs, specs) where you want persistent memory for an agent without the infrastructure overhead.

Repo: https://github.com/nonatofabio/local_faiss_mcp

I’d love feedback on the implementation—specifically if anyone has ideas on better handling the chunking logic without bloating the dependencies, or if you run into performance issues with larger indices (10k+ vectors).

Comments URL: https://news.ycombinator.com/item?id=46136319

Points: 1

# Comments: 0