I build the layer between machine-learning models and hardware they never talk to directly: GPU bridges with no native bindings, neuroevolution engines on flat typed arrays and Rust-WASM kernels, semantic search over codebases, orchestration runtimes for LLM agents.

The common thread: make the slow thing fast and the fragile thing crash-safe — then prove it with tests, not adjectives.

7 public repos · 5 packages on npm · MIT / Apache-2.0 · Bun + TypeScript strict

tfjs-turbo — GPU without the native-binding tax. tfjs-node ships a compiled .node binding that breaks across OS and Node ABI combos, and its CUDA path is Linux-only. tfjs-turbo runs TensorFlow.js inside headless Chrome and bridges training back to Node/Bun over a control channel — WebGPU first, WebGL/WASM as automatic fallback. Training survives process crashes via IndexedDB-backed checkpoint & resume, and the examples double as the CI suite.

import { TensorFlow } from 'tfjs-turbo';

const tf = new TensorFlow({ backend: 'wasm' });
await tf.ready();

const result = await tf.train({
    layers: [
        { type: 'dense', units: 64, activation: 'relu', inputShape: [10] },
        { type: 'dense', units: 1, activation: 'sigmoid' }
    ],
    compile: { optimizer: 'adam', loss: 'binaryCrossentropy' },
});

bunaptic — neuroevolution rebuilt for modern runtimes. Neataptic proved flexible JS neural-network APIs are worth having; its runtime predates typed arrays, worker threads and WASM. Bunaptic keeps the API spirit and rebuilds the engine: flat Float64Array genomes, population evaluation across Workers, Rust-WASM kernels for dot-product-heavy paths, dense feed-forward fast paths. The README states the alpha limits on purpose — recurrent adam/rmsprop still run the TypeScript BPTT path.

general.ai — orchestration as a protocol, not a wrapper. Raw SDK calls make agent behavior drift. general.ai adds a protocol layer on top of any OpenAI-compatible endpoint: tool and subagent definitions, retries, provider key rotation, request queueing, context compression and structured checkpoints — plus a native mode for when you want exact SDK semantics and nothing else.

teachable-machine.js — inference that respects your RAM. Batched classification with a RAM-first pipeline, automatic disk fallback under memory pressure, and strict cleanup guarantees so long-running services don't leak. Video inputs are frame-sampled through FFmpeg; images and videos share one ergonomic API accepting URLs, paths, buffers and data URIs.

opencode-beacon — search code by meaning. Embeddings alone miss exact identifiers; keywords alone miss intent. Beacon fuses vector + BM25 + identifier boosting into hybrid recall, then builds on it: dependency graphs, change-impact analysis ("what breaks if I touch this file?"), temporal search across git history and semantic diffs. Fifteen tools exposed to the agent.

timeline
    title From game tooling to ML infrastructure
    2025 : robify — Roblox FFlag library
         : teachable-machine.js — ML inference on Node
         : tfjs-turbo — GPU without native bindings
    2026 : general.ai — LLM orchestration runtime
         : bunaptic — neuroevolution on Bun + Rust-WASM
         : opencode-beacon — semantic tooling for AI agents

Measured, not vibes. A monthly snapshot is appended automatically (scripts/npm-metrics.sh on a systemd timer) to metrics/npm-downloads.md.

Baseline 2026-10-04: 230 downloads/month across the five packages · 8 stars total. Small numbers, published on purpose — the trend line is the point.

• Bun-first. TypeScript strict, tests colocated with source, bun test as the merge gate.
• Tests are the spec. Crash-resume, backend fallback and folding correctness are proven by suites, never claimed in prose.
• Root cause over symptom. A fixed bug ships together with the test that would have caught it.
• Honest status labels. Alpha means alpha; download counts get published as-is, drift and all.

• Email — codelabsnixaut@gmail.com
• GitHub — @nixaut-codelabs

_{npm versions and last-month download counts checked 2026-10-04. They will drift; publishing them anyway is the point.}