Skip to content
View ashvardanian's full-sized avatar
πŸ’­
Less Slow
πŸ’­
Less Slow

Sponsoring

@martinus
@jart
@danmar
@jedisct1
@nlohmann
@networkx
@python
@skypjack
@sharkdp
@ETLCPP
@maxbachmann

Highlights

  • Pro

Block or report ashvardanian

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Please don't include any personal information such as legal names or email addresses. Maximum 100 characters, markdown supported. This note will be visible to only you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ashvardanian/README.md

Hey, I'm Ash!

I've designed Search & AI infra software deployed on hundreds of millions of devices, which power some of the largest tech companies, unicorns and decacorns, AI research labs, and defense and intelligence organizations worldwide. In short:

  • CS/AI researcher.
  • Astrophysics dropout.
  • Nomad, somewhat digital.
  • Building the Unum since 2015.
  • Investing in deep-tech, cloud, & semiconductors.
  • Fluent in English, Russian, & Armenian.

I spent most of the last 20 years writing code. Mostly GPGPU, SIMD Assembly, userspace storage drivers, and networking protocols. I strongly believe, that:

  • procedural programming eats object-oriented for breakfast,
  • abstractions are mostly evil, they are never zero-cost,
  • infrastructure is fun,
  • less code is better than more code,
  • east-const is superior to const-west in C/C++,
  • goto is a blessing, not a curse,
  • good is the enemy of great – elegant code is better than working code,
  • the future of AI is amazing, and it's not LLMs or transformers,
  • matrix multiplications are tasty, but literally everyone cooks them wrong,
  • companies grow faster than people,
  • pivots are the plague of Silicon Valley,
  • hard things are easy, and easy things are hard,
  • how you do one thing - is how you do every thing,
  • talent, ambition, and dedication are much more important than material resources,
  • a single developer can beat a trillion-dollar company in almost everything.

If you want to get in touch and contribute - my handle is @ashvardanian on most platforms, including Twitter/X and LinkedIn. If you need funding for something deeply technical - ping me at ash@aloniq.com and ash@aal.vc. I value a codebase no less than a pitch deck πŸ€—

USearch Python installs UCall Python installs SimSIMD Python installs StringZilla Python installs GitHub Org's stars: unum-cloud GitHub User's stars: ashvardanian HackerNews User Karma

Repositories

I've designed and maintain the following libraries, datasets, and AI models:

  • StringZilla - faster strings with SIMD/SWAR for C, C++, Rust, Python, and Swift, replacing GlibC and STL
  • USearch - single-file search engine for C, C++, Rust, Swift, Go, Java, C#, Python, JS..., replacing FAISS
  • UForm - small multimodal pre-trained AI models with SDKs for Python, JS, and Swift, replacing CLIP
  • UCall - networking server library for C and Python designed for Linux kernel bypass, replacing FastAPI
  • SimSIMD - fast vector-vector math library for C, Python, Rust, and JavaScript, replacing BLAS level 1

Micro repositories:

  • USearch-Molecules - 28 billion embeddings - one of the largest datasets in Computational Chemistry
  • UStore - multimodal embedded database for C, C++, and Python designed around key-value stores
  • SwiftSemanticSearch - example of on-device real-time AI using UForm and USearch on iOS
  • cpp-cuda-python-starter-kit - C++ and CUDA starter kit for Python developers
  • TinySemVer - Semantic versioning GitHub CI tool that doesn't take 300K lines of JavaScript
  • affine-gaps - Less wrong local and global Gotoh sequence alignments in one NumBa Python file
  • ParallelReductionsBenchmark - GPGPU benchmarks for SyCL, CUDA, OpenCL, Vulkan, etc.
  • BenchmarkingTutorial - C/C++ tutorial for performance-oriented programming using Google Benchmark
  • memchr_vs_stringzilla - Rust micro-benchmark comparing StringZilla to the MemChr crate
  • usearch-benchmarks - Billion-scale Vector Search benchmarks against FAISS, Weaviate, Qdrant, etc.
  • ucsb - parallel benchmarks for ACID persistent key-value stores, like RocksDB
  • scaling-democracy - combinatorial voting method implemented in CUDA for H100 GPUs

Materials

Cherry picks:

Pinned Loading

  1. unum-cloud/usearch unum-cloud/usearch Public

    Fast Open-Source Search & Clustering engine Γ— for Vectors & πŸ”œ Strings Γ— in C++, C, Python, JavaScript, Rust, Java, Objective-C, Swift, C#, GoLang, and Wolfram πŸ”

    C++ 2.3k 147

  2. StringZilla StringZilla Public

    Up to 10x faster strings for C, C++, Python, Rust, and Swift, leveraging NEON, AVX2, AVX-512, and SWAR to accelerate search, sort, edit distances, alignment scores, etc πŸ¦–

    C++ 2.3k 80

  3. unum-cloud/ucall unum-cloud/ucall Public

    Web Serving and Remote Procedure Calls at 50x lower latency and 70x higher bandwidth than FastAPI, implementing JSON-RPC & REST over io_uring ☎️

    C 1.2k 43

  4. SimSIMD SimSIMD Public

    Up to 200x Faster Dot Products & Similarity Metrics β€” for Python, Rust, C, JS, and Swift, supporting f64, f32, f16 real & complex, i8, and bit vectors using SIMD for both AVX2, AVX-512, NEON, SVE, …

    C 1.1k 64

  5. unum-cloud/uform unum-cloud/uform Public

    Pocket-Sized Multimodal AI for content understanding and generation across multilingual texts, images, and πŸ”œ video, up to 5x faster than OpenAI CLIP and LLaVA πŸ–ΌοΈ & πŸ–‹οΈ

    Python 1.1k 62

  6. unum-cloud/ustore unum-cloud/ustore Public

    Multi-Modal Database replacing MongoDB, Neo4J, and Elastic with 1 faster ACID solution, with NetworkX and Pandas interfaces, and bindings for C 99, C++ 17, Python 3, Java, GoLang πŸ—„οΈ

    C++ 554 31