Projects

The first Apache Arrow project to run compute on the Apple silicon GPU. Every Mac since 2020 shares one physical memory between CPU and GPU, so an Arrow buffer in Metal shared memory is already a GPU buffer, with nothing to upload or download. ArrowMetal runs 307 of Arrow's 307 compute functions there, honouring validity bitmaps, offsets, dictionaries and nested types, from Python, Rust, Swift, Go, R, TypeScript and C over one C ABI. Measured, not claimed: 339 benchmark rows on an Apple M4 Max (16 cores, 50M rows) against the fastest CPU idiom of Polars, pyarrow, pandas and numpy — take at 24.2× pyarrow on 25M random indices, string contains 7.0× Polars lazy on 10M rows, sort float64 4.1× Polars, group-by sum 3.8× pyarrow Acero — and the 77 rows where the CPU is still ahead published next to them, grouped by cause. 39,069 differential test cases against pyarrow; fourteen findings reported upstream, two open as Apache Arrow pull requests. v0.1.0 on PyPI and crates.io; an independent Apache-2.0 project, listed on Apache Arrow's Powered By page.

An Apache Arrow ADBC driver for any ODBC data source. One plain-C11 shared library that turns every ODBC driver on your machine into an Arrow-native ADBC driver — columnar record batches out, bulk ingest in — from Python, Rust, Go, Java and C#. Measured, not claimed: 46 databases verified on Linux, 41 on macOS (Apple Silicon), 45 on Windows through one workload, five language bindings measured against all of them on one binary, and every number named with the laptop and the load it was taken under. Reads land within 7% of the raw ODBC floor and 2.4× ahead of pyodbc; partitioned reads beat the native PostgreSQL ADBC driver at eight connections on a quiet host, while a single connection stays at 0.4× and bulk ingest never clears native COPY — both stated up front. Where a native ADBC driver exists it hands the connection over. v0.1.0 on PyPI, crates.io and NuGet; an independent Apache-2.0 project that implements the ADBC standard, listed on Apache Arrow ADBC's official Tools & Integrations documentation page.

A tiny, dependency-light HNSW vector index for approximate nearest-neighbor search — built to compile to WebAssembly. serde + postcard are its only Rust dependencies, and the wasm build is reported at ~33 KB gzipped. unsafe is denied crate-wide outside one audited SIMD kernel, and a deterministic splitmix64 PRNG keeps getrandom out of the tree, so it lands on wasm32-unknown-unknown with no shims. The npm package layers transformers.js auto-embedding on a Web Worker, OPFS persistence, and cross-tab single-writer leader election — a three-line semantic-search API that runs entirely in the browser, no server. v0.3.1 on crates.io and npm.

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

GPU-accelerated DuckDB extension for NVIDIA CUDA and Apple Silicon Metal — the first SQL execution engine to target Apple Silicon GPUs, and now an official DuckDB Community Extension: INSTALL gpudb FROM community in any DuckDB ≥ 1.5.5, signed, with the full Metal backend on M-series Macs. v0.3.0's streaming-aggregate rewrite reaches native-DuckDB parity end-to-end on TPC-H (the worst cell improved ~100×), while the Metal operator kernels keep 3–25× wins (1187 GiB/s on RTX 4090, ~470 GiB/s on M4 Max). A community-contributed 9.9× Metal hash join lands next.

A code editor built for the AI era. Native C++/Rust, a tiny no-Electron binary (~12 MB, ~4.4 MB compressed on Linux), runs on Linux/macOS/Windows. Memory-mapped I/O opens files up to ~1.8GB without loading them into RAM. 238 file types, 82 language lexers. Local-first by design; cloud LLMs are opt-in. No telemetry, no analytics, no mandatory API key.

The lightweight screenshot tool for every grid. Native Rust app for Linux/macOS/Windows — same hotkeys, overlay, and nine annotation tools everywhere. No Electron, no telemetry. Capture regions, annotate, save, pin to screen.

Fast, beautiful disk space analysis for Linux, macOS, and Windows. A single ~6 MB native Rust binary — no Electron, no webview — that scans every CPU core in parallel and shows where your space went three ways: cushion treemap, heatmap grid, and radial sunburst. Cross-linked dashboard (tree ↔ map ↔ file types ↔ largest files), in-place delete with no rescan. Private by design: binds to 127.0.0.1 only, no telemetry, nothing ever leaves your machine. And because a scan is the perfect moment to goof off, every build ships with The Seeker's Voyage — a fully offline mini-game you sail through while Diskhoji scans in the background, on all three platforms.

Every byte reaches the far shore. A fast, resumable file copier and mover for large folders — or a single big file — onto external drives. Named for the boatman: it survives unplugs, crashes, and sleep, then resumes exactly from the last proven checkpoint instead of starting over. One small native Rust binary that's both a window and a CLI — run it bare and it opens a CPU-rendered window (no OpenGL); give it two paths and it stays a command-line tool. ~5.5 MB app, or a ~0.7 MB static musl CLI-only build for servers and containers. Installers for Windows (MSI), macOS (DMG), and Linux (AppImage); every release ships SHA256SUMS covering every asset. No network, ever. MIT, on crates.io.

A production-grade Rust SDK for Anthropic's Claude API. Every wire enum carries an Unknown catch-all that preserves unfamiliar JSON and re-serializes it verbatim, so a response from a model the SDK has never heard of round-trips without an SDK update. It type-checks clean for wasm32-unknown-unknown, a target where tokio leaves the dependency graph entirely (on native it arrives transitively through reqwest). 191 tests pass, 0 fail. An independent project, not affiliated with Anthropic.

Free, browser-based AI education for everyone — in English & Hindi, no app, no fee, free forever. 17 courses from the basics to advanced, about ten minutes a day, anchored by two flagship tracks: AI ki ABCD (AI fundamentals + practical use) and Job-Ready Basics (digital skills + interview prep). An AI Lesson Helper answers your questions from the lesson content in Hindi or English. Made for whoever you are — new starters, students, working professionals, and curious explorers.