PS

Prateek Singh

Researcher, builder, and storyteller.

Built ArrowMetal — listed on Apache Arrow's official Powered By page, the first Apache Arrow project to run compute on the Apple silicon GPU. An Arrow buffer in Metal shared memory is already a GPU buffer, so ArrowMetal runs 307 of Arrow's 307 compute functions there, from seven languages: take at 24.2× pyarrow on 25M indices and sort 4.1× Polars on 50M rows, on an M4 Max, with all 339 measured rows published.

Built adbcBridge — an Apache Arrow ADBC driver for any ODBC data source, listed on Apache Arrow ADBC's official Tools & Integrations page: one plain-C11 library that turns every ODBC driver on your machine into an Arrow-native one, from Python, Rust, Go, Java and C#. 46 databases verified on Linux, 41 on macOS, 45 on Windows — every number measured on an ordinary laptop and named with its load.

Built gpudb — the first GPU SQL engine for Apple Silicon, now an official DuckDB Community Extension: one line installs Metal kernels into any DuckDB ≥ 1.5.5, with native-parity TPC-H end-to-end and 3–25× operator wins on an M4 Max.

Wrote The First Mind — a 20-chapter illustrated novel that tries to make learning what an LLM actually is genuinely fun. Five friends, one weekend hackathon, a year of late nights. Comic panels for every scene, audio narration in English and Hindi.

Launched Skill the Gap — free, browser-based and mobile-friendly AI courses for everyone, in English & Hindi. Seventeen courses from the basics to advanced, about ten minutes a day, free forever — no app, no fee, made for whoever you are.

Also four small native Rust tools — Notepatra (a code editor for the AI era), KAShot (a cross-platform screenshot tool), Diskhoji (a disk-space analyzer with a mini-game to play while it scans), and Kevat (a resumable file copier that survives unplugs and picks up from the last proven checkpoint). All four tiny native binaries — no Electron, no telemetry — all on the Projects page.

Published crimson-crab — a production-grade Rust SDK for Anthropic’s Claude API: streaming, tool use, extended thinking, prompt caching, and batches. 191 tests, a panic-free library enforced at compile time, a tokio-free public API, and a clean wasm32 type-check. v0.1.0, MSRV 1.75, dual MIT OR Apache-2.0. An independent project, not affiliated with Anthropic.

Published ferrovec — a tiny, dependency-light HNSW vector index built to compile to WebAssembly. serde & postcard are its only Rust dependencies, the wasm build is reported at ~33 KB gzipped, and a deterministic PRNG keeps getrandom out of the tree so it lands on wasm32 with no shims. The npm package adds transformers.js auto-embedding on a Web Worker, OPFS persistence, and cross-tab leader election — semantic search that runs entirely in the browser, no server. v0.3.1 on crates.io & npm.

The throughline behind the teaching: make AI genuinely accessible to everyone, for free, in their own language — a free course in English & Hindi, and a story that teaches what an LLM actually is.

theaivibe.org publishes all of itOpen sourceFree to read

Teaching AI to everyone — for free

The same idea, two forms: free AI education in your own language. One teaches the skills, the other teaches the ideas — both in English & Hindi, both free forever.

Projects

Ten shipped, all in active development — nine open-source projects on GitHub, plus Skill the Gap, a free AI-learning platform for everyone.

ArrowMetal

The first Apache Arrow project to run compute on the Apple silicon GPU. 307 of Arrow's 307 compute functions from Python, Rust, Swift, Go, R, TypeScript and C; take at 24.2× pyarrow on 25M indices and sort 4.1× Polars on 50M rows, on an M4 Max; 339 rows measured against the fastest CPU idiom. Independent Apache-2.0 project, listed on Apache Arrow's official Powered By page. v0.1.0 on PyPI and crates.io.

adbcBridge

An Apache Arrow ADBC driver for any ODBC data source — one plain-C11 library that turns every ODBC driver into an Arrow-native one, from Python, Rust, Go, Java and C#. 46 databases verified on Linux, 41 on macOS, 45 on Windows; every number measured. Listed on Apache Arrow ADBC's official Tools & Integrations page. v0.1.0 on PyPI, crates.io and NuGet.

ferrovec

A tiny, dependency-light HNSW vector index built to compile to WebAssembly. serde + postcard the only Rust deps, ~33 KB gzipped wasm, no unsafe outside one SIMD kernel. transformers.js auto-embedding, OPFS persistence, cross-tab leader election — semantic search entirely in the browser. v0.3.1 on crates.io + npm.

Turning Dials

Interactive AI education at distrum.app — learn AI from the smallest possible piece. Live, runnable simulations in the browser: pause a neural network mid-training, turn the dials, and watch real learning happen, from the first neuron up through transformers and RL. Built to cost a student nothing.

gpudb

The first GPU SQL engine for Apple Silicon — now an official DuckDB Community Extension. INSTALL gpudb FROM community: native-parity TPC-H end-to-end, 3–25× Metal operator kernels, CUDA on Linux.

Notepatra

Code editor built for the AI era. Native C++/Rust, a tiny no-Electron binary, all 3 platforms. Local-first AI, no telemetry.

KAShot

Lightweight screenshot tool for every platform. Native Rust, no Electron, no telemetry, 9 annotation tools.

Diskhoji

Disk space analyzer for every platform. One ~6 MB native Rust binary — treemap, heatmap & radial views over a parallel scanner. Plus a built-in mini-game to sail while it scans. No telemetry, private by design.

Kevat

Resumable file copier & mover for large folders onto external drives — survives unplugs and crashes, resumes from the last proven checkpoint. One small native Rust binary, window + CLI in one, ~0.7 MB static CLI build for servers. No network, ever. MIT.

crimson-crab

Production-grade Rust SDK for Anthropic's Claude API — streaming, tool use, extended thinking, prompt caching, batches. 191 tests, panic-free at compile time, a tokio-free public API, and a clean wasm32 type-check. An independent project, not affiliated with Anthropic.

Skill the Gap

Free AI learning for everyone — 17 courses from the basics to advanced, in English & Hindi. Playful, jargon-free, ten minutes a day. No app, no fee, free forever. Made for whoever you are.

Currently shipping

  • ArrowMetal v0.1.0 — just shipped: the first public release is on PyPI and crates.io (pip install arrowmetal), and it made it onto Apache Arrow's official Powered By page. Next up: pipelining for the dispatch floor; benchmark files from Macs that are not an M4 Max; and the upstream reports, two of them open as Apache Arrow pull requests. Read the launch post.
  • adbcBridge v0.1.0 — just shipped: the first public release is on PyPI, crates.io and NuGet (pip install adbcbridge). Next up: the Win32 thread shim that restores prefetch and parallel ingest on Windows, a driver bootstrap for the open-licence ODBC drivers, Maven Central, then a JDBC bridge on the same model. The launch post has every number with its host and load attached.
  • gpudb v0.4 — v0.3.0 now ships signed from the DuckDB Community Extensions registry (INSTALL gpudb FROM community, one line, no manual download). Next up: landing the community-contributed Metal hash join — verified 9.9× on a 1M × 10M inner join — plus GPU window operators on the road to v1.0.
  • Podcast distribution — Apple Podcasts + Spotify submission for The First Mind audio narration. The iTunes-format RSS feed is built; just needs to land in the directories.

Get in touch

Open to collaboration on research, GPU systems, free AI education, or the stories.