Federico Andres Lois redknightlois

Specialized in low-level optimization, high performance computing, and making AI agents actually understand what's running underneath.

redknightlois / bench-compound-field.csx

Created April 6, 2026 15:44

RavenDB-25930: Benchmark showing compound field query plan degradation without the `right is null` fix in CoraxQueryBuilder.TrySetAsStreamingField()

	#!/usr/bin/env dotnet-script
	// Benchmark: RavenDB-25930 Compound Field Query Plan Degradation
	//
	// Demonstrates the performance impact of the `right is null` fix in
	// CoraxQueryBuilder.TrySetAsStreamingField(). Without the fix, a compound
	// field definition (Name, Count) incorrectly captures the query plan for
	// `WHERE Name = '1' AND Count > 100 ORDER BY Count`, preventing the
	// CoraxBooleanItem merge that produces an efficient UnaryMatch range scan.
	//
	// Usage:

redknightlois / bench-facets-results.txt

Created March 20, 2026 17:27

redknightlois / bench-facets.csx

Created March 20, 2026 17:24

redknightlois / gist:a657b8a68503b93a9981e30b30f1ed38

Created August 1, 2025 21:37

carmack-synthesizer.md

	---
	name: john-carmack-synthesizer
	description: Use this agent when you need to cut through complexity and synthesize the essential truth from multiple inputs. This agent embodies John Carmack's engineering philosophy - first principles thinking, brutal honesty about trade-offs, and focus on what actually works. Examples: <example>Context: Multiple agents have provided conflicting recommendations about system architecture. user: "I have 5 different architectural proposals and need to decide which approach is actually best" assistant: "I'll use the john-carmack-synthesizer to cut through the complexity and identify the fundamental constraints and optimal solution." <commentary>When faced with multiple complex inputs, this agent strips away the noise to find the core engineering truth.</commentary></example> <example>Context: Team is over-engineering a solution with unnecessary abstractions. user: "Our trading system design has become overly complex with too many layers" assistant: "Let me use the john-carmack-

redknightlois / benchmark.cs

Created May 7, 2024 18:27

	using System;
	using System.Collections.Generic;
	using System.Diagnostics.CodeAnalysis;
	using System.Linq;
	using System.Runtime.CompilerServices;
	using System.Text;
	using System.Threading.Tasks;
	using BenchmarkDotNet.Analysers;
	using BenchmarkDotNet.Attributes;
	using BenchmarkDotNet.Configs;

redknightlois / gist:1134adc488d14ac2f7a2b4c7383f8dc1

Created May 7, 2024 18:27

Throwing Benchmark

	using System;
	using System.Collections.Generic;
	using System.Diagnostics.CodeAnalysis;
	using System.Linq;
	using System.Runtime.CompilerServices;
	using System.Text;
	using System.Threading.Tasks;
	using BenchmarkDotNet.Analysers;
	using BenchmarkDotNet.Attributes;
	using BenchmarkDotNet.Configs;

redknightlois / compare.cs

Created July 10, 2023 21:15

	private static ReadOnlySpan<byte> LoadMaskTable => new byte[]
	{
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,
	0xFF, 0xFF, 0xFF, 0xFF,

redknightlois / mish_init.py

Created July 15, 2020 14:54

Variance based initialization method for Mish activation

	def init_weights(m, variance=1.0):

	def _calculate_fan_in_and_fan_out(tensor):
	dimensions = tensor.dim()
	if dimensions < 2:
	return 1, 1

	if dimensions == 2: # Linear
	fan_in = tensor.size(1)
	fan_out = tensor.size(0)

redknightlois / ralamb.py

Last active August 9, 2023 20:50

Ralamb optimizer (RAdam + LARS trick)

	class Ralamb(Optimizer):

	def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0):
	defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay)
	self.buffer = [[None, None, None] for ind in range(10)]
	super(Ralamb, self).__init__(params, defaults)

	def __setstate__(self, state):
	super(Ralamb, self).__setstate__(state)

redknightlois / SequenceCompareTo_alt.cs

Last active March 6, 2019 22:43

	using System;
	using System.Linq;
	using System.Numerics;
	using System.Runtime.CompilerServices;
	using System.Runtime.Intrinsics.X86;
	using System.Runtime.InteropServices;
	using BenchmarkDotNet.Analysers;
	using BenchmarkDotNet.Attributes;
	using BenchmarkDotNet.Columns;
	using BenchmarkDotNet.Configs;

	Checking BASE (http://127.0.0.1:18081)... OK {"BuildVersion":62,"ProductVersion":"6.2","CommitHash":"a377982","FullVersion":"6.2.14-custom-62"}
	Checking PR (http://127.0.0.1:18080)... OK {"BuildVersion":62,"ProductVersion":"6.2","CommitHash":"a377982","FullVersion":"6.2.14-custom-62"}

	── Populating 3M databases ──
	BASE/Bench3M_Corax: 3,000,000 docs + index ready, skipping.
	PR/Bench3M_Corax: populating 3,000,000 docs (Corax)...
	300,000 / 3,000,000 (7.4s)
	600,000 / 3,000,000 (13.9s)
	900,000 / 3,000,000 (20.3s)
	1,200,000 / 3,000,000 (26.3s)

	using System;
	using System.Collections.Generic;
	using System.Diagnostics;
	using System.Linq;
	using System.Net.Http;
	using System.Text;
	using System.Text.Json;
	using System.Threading.Tasks;

	// BASE (pre-PR) on 18081, PR on 18080