Micha(el) Bladowski michabbb

CLM-v0.1-8B vs. JEV as a relatedness judge: a small, reproducible field test

TL;DR: On 140 real entity pairs from a production knowledge-graph linking pipeline, the open CLM-v0.1-8B (Contrastive-LM, with its official encoder Qwen3-8B BF16 on vLLM) could not replace JEV (typesafe-ai/jev, hosted) as a yes/no "are these two entities related?" judge:

	CLM (last 2048 tokens, default)	CLM (full input)	JEV
AUC vs. 40 gold labels	0.77	0.64	0.998
Output range of P(related)	0.22 – 0.77	0.49 – 0.64	0.07 – 0.86

PHP-FPM: Route a long-running endpoint into a separate pool — without touching application code

Problem

PHP-FPM has three different timeouts that can kill a request:

Layer	Directive	Overridable per-request from PHP?
nginx	`fastcgi_read_timeout`	❌ No (the connection is dropped on the nginx side)
PHP	`max_execution_time` (php.ini)	✅ Yes, via `set_time_limit()` / `ini_set()`

MedGemma MRI Analysis Script

A Python script for analyzing MRI DICOM images using Google's MedGemma model via Vertex AI.

DISCLAIMER: This is for educational/experimental purposes only. NEVER rely on AI for medical diagnosis!

What This Does

Sends batches of DICOM images to MedGemma for analysis
Handles the tricky Vertex AI endpoint configuration

Building an OpenCode CLI Plugin for Audio Permission Notifications

Overview

This guide shows how to create an OpenCode CLI plugin that plays an audio notification when OpenCode needs user permissions. The solution consists of two components:

A shell script to play system audio
An OpenCode plugin that listens for the permission.asked event

The Great LLM Code Optimization Battle: 8 AI Models, 1 Function

The Challenge

What happens when you pit the world's most advanced AI models against each other in a code optimization challenge? I decided to find out by asking 8 different large language models to optimize a single PHP function - and the results were fascinating, surprising, and sometimes hilariously broken.

The prompt was simple and identical for all contestants:

"optimize this code: php 7.4"

SSH Post-Quantum Cryptography Configuration Guide

This guide explains how to configure SSH to use post-quantum cryptography and suppress related warnings.

Understanding the Warning

When connecting via SSH, you might see:

** WARNING: connection is not using a post-quantum key exchange algorithm.
** This session may be vulnerable to "store now, decrypt later" attacks.

	Open-source alternatives
	1. OPENjutsu (D3voz): replaces Genjutsu, runs locally
	https://github.com/D3voz/openjutsu
	2. Alter (Scenario): replaces Genjutsu, needs the cloud
	https://www.scenario.com/explorations/alter/
	3. Sirio Berati's Workflow: replaces Genjutsu, uses a paid API
	https://github.com/sirioberati/Genjustsu-Open-Source-Workflow
	4. Freefield (neuregex): recreates 120+ Higgsfield apps
	https://github.com/neuregex/freefield
	5. Open Source Canvas (Raphael Liu): replaces Canvas

	---
	name: self-reminder
	description: Schedule Codex reminders and self-wakeups through cron and HerdR. Use for “remind me in …”, “check every X minutes/hours …”, “wake yourself up in …”, and requests to list or cancel these reminders. Supports one-time reminders and recurring checks of ongoing work.
	metadata:
	targets: [codex]
	---

	# Self-reminder

	Use cron to send a future message to this Codex session through HerdR, then yield the turn. This allows background work to continue without keeping the agent in a polling loop.

	#!/usr/bin/env python3
	"""Cron-backed messages to the originating Codex session in HerdR."""
	import argparse
	import fcntl
	import json
	import os
	from pathlib import Path
	import re
	import shlex
	import subprocess

	#!/bin/bash
	#
	# check_pve_backup - Nagios/Icinga plugin for Proxmox VE backup monitoring
	#
	# Monitors vzdump backup status via the Proxmox API (pvesh).
	# Supports checking specific VMs, all running VMs, or any recent backup.
	#
	# Requirements:
	# - Must run on a Proxmox VE host (needs pvesh command)
	# - jq for JSON parsing