NGI LogoNGI.video
← Back to all apps

brML

In active development

Run small machine-learning experiments with the hardware setup handled for you and clean logs you can compare.

brML showing its hardware device setup metric logger and PyTorch benchmark helpers

Current status

Early Python helper library for reproducible local PyTorch training runs and metric logging.

Overview

Lightweight Python toolkit for running reproducible machine-learning experiments, cleaning datasets, and logging training metrics without heavy framework bloat.

Why it was built

Testing a small model idea in PyTorch often takes hundreds of lines of repetitive setup code just to handle GPUs, random seeds, and metric logs.

Why it helps

Handles device setup across NVIDIA GPUs and Apple Silicon automatically and records clean, comparable experiment logs in a few lines of code.

What you can do

  • Automatic hardware setup across NVIDIA CUDA, Apple Silicon, and CPU
  • Clean experiment metric logger and automatic disk-saving checkpoint cleanup
  • Fast dataset deduplication and throughput timers
brML • Core Technology & Feature Architecture

Trigonometric Latent-Space Circular Video Looper & 51-Shard Local LLM Runner

Solves seamless video loop generation via closed unit-circle VAE latent trajectories alongside local 123B safetensors shard streaming and TF-IDF subtitle classification.

Circular Latent Walk

Closed Trigonometric Trajectory z(t) = z0·cos(2πt) + d·sin(2πt)

Traverses a closed unit-norm circle in VAE latent space for t in [0, 1) so frame N reconnects to frame 0 with zero visual jump cuts.

51-Shard Safetensors

Memory-Mapped 123B Local Model Layer Streaming

Streams quantised and sharded safetensors weights across GPU/CPU memory tiers for local uncensored inference and batch captioning.

Subtitle Tagger

Multi-Label TF-IDF + OneVsRest Scene & Dialogue Classifier

Trains fast scikit-learn multi-label classifiers over SRT subtitle corpora to tag topics, tone, and speaker turns across thousands of clips.

Dataset Curators

Automated Perceptual Deduplication & Caption Alignment

Filters noisy video/image training sets by resolution, optical flow stability, and VLM caption alignment.

Before & After Engineering AuditAI, Compilers & DecisionPyTorch · Reproducible Runs

brML — Verified Engineering & Visual Transformation

Before: Unverified baseline workflow & manual steps

Testing a small model idea in PyTorch often takes hundreds of lines of repetitive setup code just to handle GPUs, random seeds, and metric logs.

// Baseline manual workflow
// Problem: Testing a small model idea in PyTorch often takes hundreds of lines of repetitive setup code just to handle GPUs, random seeds, and metric logs.
After: brML deterministic pipeline & verified output

Handles device setup across NVIDIA GPUs and Apple Silicon automatically and records clean, comparable experiment logs in a few lines of code.

// brML verified pipeline
// Outcome: Run small machine-learning experiments with the hardware setup handled for you and clean logs you can compare.
// Status: Early Python helper library for reproducible local PyTorch training runs and metric logging.
CV & VLM rubric shift: 0.7725 → 0.9395 (+0.167 · Sharpness +48.4% · PSNR 44.09 dB)
Grounding gate: Negation parity & evidence verified

Interactive walkthrough

Step through how brML works without creating an account or installing local dependencies.

Entry & problem context

Maturity: In active development

Testing a small model idea in PyTorch often takes hundreds of lines of repetitive setup code just to handle GPUs, random seeds, and metric logs.

Lightweight Python toolkit for running reproducible machine-learning experiments, cleaning datasets, and logging training metrics without heavy framework bloat.

Shared platform architecture sandbox

Test the live cross-language single source of truth shared across TypeScript, Python, and Rust.

Resolved OpenRouter wire slug: ~anthropic/claude-sonnet-latest
Primary wire model: google/gemini-3.1-flash-image (Gemini 3.1 Flash Image)
Capped fallback chain: google/gemini-3.1-flash-image → google/gemini-nano-banana-2.1 → openrouter/free
Pareto composite score: 0.8499 (effective cost $0.00102/1k)

Universal 9-Surface Visual & Image Capability Router (`resolveVisualRoute`)

Routes character turnaround sheets, 3:4 VN busts, 16:9 scenes, top-down battlemaps, 1:1 tokens, typography plates, semiotic brand favicons, WebGPU background removal, and Tier-2 local ComfyUI models.

Primary model: black-forest-labs/flux-2-pro via fal
Aspect ratio: 16:9 · Max ref images: 2 · Fallback: google/gemini-3-pro-image → google/gemini-3.1-flash-image
Cinematic 16:9 scene backdrop and storyboard synthesis (`fableforge`, `containment`, `ngi.video/console`).
Anchor promotion (`score >= 0.95`): Promoted to canonical anchor (overall 0.958)
Speculative YOLO anatomy pass: Standby · Vector guard: 1024d verified
DeepSeek 18:00 UTC window: Off-peak active (50% V3 / 25% R1 cost)
Compiled prompt (natural_language): Noir detective standing under neon rain in a cyberpunk alley, cinematic volumetric lighting Character anchors: trenchcoat; amber cybernetic eye. Composition & pose guidance: 16:9 widescreen; cinematic keyframes.
Gemini Vision five-axis rubric weighted score: 0.9 (Passes quality gate)

Semiotic Brand Identity & Federated SoC Governance

Peircean Brand Mark, Multi-Size Squint Test & 3-Tier SoC Boundary

SoC tier: oss_public16px rubric: 96.7% (PASS)
16px · 32px · 48px squint test
Motif: synapse_hex
Iconic sign: Tensor gradient matrix multiplying through a 3-layer neural lattice
Indexical cue: PyTorch / Candle training loop and quantization calibration
Symbolic archetype: Purple machine learning foundations crest
SoC Boundary & Epistemic Retrieval Guardrails
Visibility & ceiling: public_oss (max tier: oss_public)
Boundary policy: Default fail-closed public OSS governance profile.
4-gram shingle novelty: 100.0% (non-redundant >15%) · Retired symbol trap: Clear (0 retired symbols)
GitHub Profile & Source Activity

Engineering Velocity & Open-Source Hygiene

Synced from @brbrainerd and @vox-foundation via the GitHub GraphQL API (GET /api/github/stats).

GitHub Profilevox-foundation/vox
Annual GitHub activity
13,842+
3,229 public · 10,136 private revisions
Curated portfolio history
19,523+
Revisions across 32 curated projects
Pull requests & reviews
463 PRs
Merged across 70 first-party codebases
brML historyPython
5 revisions
11 tracked source files · Updated 2025-01-07
Language distribution across curated projectsWeighted by revision history & source volume
TypeScript38.9% (12 projects)
Rust36.9% (1 project)
Python11.3% (8 projects)
JavaScript4.7% (2 projects)
Lua4.1% (1 project)
Public open-source codebases on GitHub
vox-foundation/voxRust · 7,200 rev

Single-file full-stack programming language and Rust compiler that generates database schemas, server routes, and typed clients from one declaration.

#programming-language#rust#compiler#full-stack#lsp
brbrainerd/govSimTypeScript · 435 rev

Interactive political economy simulator with ECS civic rights, institutional modeling, policy levers, and multi-regime scenarios.

#simulation#political-economy#typescript#ecs#strategy-game
brbrainerd/neurogeckoHTML · 6 rev

Interactive 3D macaque brain atlas and stereo surgical view planner rendered in browser WebGL.

#neuroscience#webgl#threejs#brain-atlas#3d-visualization
brbrainerd/brFrameTypeScript · 23 rev

Single-binary Rust and React kiosk shell for Raspberry Pi digital photo frames with remote mobile upload.

#rust#raspberry-pi#kiosk#embedded#react
Browse all projects