I'm an information engineering student in Japan building local AI systems, LLM inference tools, coding-agent workflows, and apps.
My projects usually start with a question: Can this actually work? I enjoy testing ideas on everything from Apple Silicon to aging GPUs, then publishing implementations, measurements, and their limitations.
- ๐ง Persistent/distributed cognition, memory, and AI agent architectures
- โก Local LLM inference, speculative decoding, and multi-GPU performance
- โป๏ธ Legacy-GPU software, OpenCL 1.1, and salvaged computing
- ๐ฏ๐ต Japanese tokenizers, fiction-oriented language models, and long-context evaluation
- ๐งช Reproducible benchmarks, regression tests, experiment records, and honest failure analysis
- ๐ ๏ธ macOS / Android / web apps, creative software, and developer tools
Languages: Rust ยท Python ยท Swift ยท Kotlin / Java ยท C / C++ ยท C# ยท TypeScript / JavaScript
ๆฅๆฌ่ชใงใๆดปๅใใฆใใพใใ**ใๅใใจใใใพใงไฝใใๅฎๆธฌใใๅ็พใงใใๅฝขใงๅ ฌ้ใใใ**ใใจใๅคงๅใซใใฆใใพใใ
October 2026 ยท public projects and experiments
| Area | Project / work |
|---|---|
| ๐งฌ Kamimusuhi / ใใฟใใใณ | Experimental architecture for a persistent artificial cognitive individual: continuity, provenance, memory, and delegation across models and machines. |
| โป๏ธ SalvagedComputing / ScrapGpuLibrary | Finding useful roles for obsolete GPUs and constrained machines. Working on legacy OpenCL 1.1 inference through CL11 and portable runtime components. |
| โก Strata experiments / Qwen3.8 MTP | Practical local inference with limited hardware: decoding speed, memory, quantization, and model/runtime tuning. |
| ๐ YLSB / llama-split-bench | Reproducible benchmarking of complete local-LLM configurations, including GPUs, split mode, context, KV cache, quality, and speed. |
| ๐ฏ๐ต J-series Tokenizers v2 / novllm | Japanese tokenizer design and model training/evaluation for long-form and fiction generation, with public research tooling. |
| ๐ค rustdsh / rdsh | Contributing to Rust tooling for AI coding agents: provider integrations, execution workflows, developer experience, and safety. |
| ๐ฝ๏ธ OISINT | An evidence-backed restaurant research and comparison app, including source links, uncertainty, and collaborative decisions. |
Additional experiments: NavSpec (structured naval-spec generation and validation) and Open Jev Core ML (experimental Apple Silicon inference, with documented limitations).
I founded and help run the ๅฐๅ ใฎLLM community on local-AI tooling, including ScrapGpuLibrary, rustdsh, and the community Bench of Us reports.
I also contribute to J-LLM on Hugging Face, sharing Japanese models, tokenizers, and other experimental artifacts. The J-series v2 collection covers multiple tokenizer vocabulary sizes.
Projects vary in maturity. A prototype, a benchmark, or a successful test is not a production guarantee.
- ๐ธ Ourin / ๆก้ด โ macOS-native Ukagaka-compatible baseware, built with Swift and native SHIORI runtimes.
- ๐ DiffLoupe โ a macOS-native folder/file diff tool inspired by WinMerge.
- โ๏ธ Singleline JP Font โ Japanese centerline fonts and 1-bit assets; the expanded v2 family is still a release candidate with proofreading in progress.
- โ NavSpec โ structured fictional 9089 naval specification generation, constraints, and regression evaluation.
- ๐ง Open Jev Core ML โ experimental Core ML / Apple Silicon conversion and benchmark results, including negative findings.
- ๐ ๏ธ HOI4 Modding AI Agent Skills โ reusable references and agent skills for game modding.
- ๐ฑ WattTime โ a published Android microwave heating-time converter built with Kotlin.
See more in my repositories and on Furin Lab.
idea / problem
โ
survey & design
โ
prototype
โ
measurement / benchmark
โ
failure analysis
โ
reproducible artifact
โ
next experiment
I use AI coding agents heavily, but try to keep claims separate from measurements: CI, test results, raw outputs, hashes, experiment manifests, and failure records matter more than "it seems to work".
I also run local experiments across Apple Silicon, CUDA, and legacy OpenCL systems, including mixed-generation NVIDIA GPUs, so hardware constraints are treated as part of the research rather than an afterthought.
Languages
Development / ML
PyTorch ยท Transformers ยท MLX / MPS ยท llama.cpp / ggml ยท CUDA ยท OpenCL ยท Core ML ยท Metal
GitHub Actions ยท SQLite ยท Supabase ยท Cloudflare
AI-assisted development
| ๐ Personal site | eightman999.com โ works and notes |
| ๐ Furin Lab | furinlab.com โ apps and public releases |
| ๐ค Hugging Face | Furinlab ยท J-LLM |
| ๐ฏ๐ต Japanese tokenizers | J-series v2 collection |
| ๐ค Local LLM community | jimoto-no-llm ยท Bench of Us |
| ๐ฝ๏ธ OISINT | oisint.com |
| ๐ฆ X | @eightman999 |
| ๐ฌ Discord | _eightman_ |
| ๐ณ Off-keyboard | Cooking and eating good food |
๐ฌ ๅใฏใๅบใใใฎๅบใใใใฎใ
๐ฌ ใใใฃใฆใฟใใใใจใใฉใใ้ง็ฎใ ใ
โ ็พฝไปใใจๅญใ็พฝไปใใจๅญ่ไฝ้ใ