AI systems · LLM inference · CUDA · agents

AI that runs fast, and shows its evidence.

I'm George Jiri Spitalsky. I make large language models run on hardware they were not supposed to fit, write the CUDA kernels that make them faster, and build AI applications whose answers can be traced to a source. Shipping transformer systems since 2020.

See the work Get in touch

model: 27B ternarygpu: 1 × RTX 5090api: openai-compatible

ternary_weights · {−1, 0, +1} · packed

MODEL
27B ternary

DECODE
up to 284 tok/s

GPU MEMORY
25.6–27.9 GiB

context
98,304 tokens

# one GPU, OpenAI-compatible endpoint served by vLLM
client.chat.completions.create(
model="Bonsai-2-27B", max_tokens=2048, temperature=0)
# → 246–284 decode tokens/s across restarts

tokens/s peak · 27B LLM · one GPU

model run on a 32 GB GPU

of 492 · DragonByte qualification 2026

2020

shipping transformers since

about

A lab of one. AI engineering with the numbers attached.

George (spital) is a senior C++ and Python engineer who works on the part of AI where models meet hardware. In 2026 he served a 27-billion-parameter ternary language model at up to 284 decode tokens per second on a single RTX 5090, using custom CUDA Tensor Core kernels, and ran the text portion of a 95 GB, 128B checkpoint on a 32 GB GPU by streaming weights from host memory at PCIe speed.

He also builds the applications around the models: a research workspace with hybrid retrieval and inspectable citations, live speech translation in the browser, and a secure OAuth relay for MCP agent tools. He started early: a multilingual GPT-2 chat and translation app in 2020, two years before ChatGPT, and one of the first public Czech GPT-2 models.

Much of his recent work is built with coding agents, with several models reviewing each other and every change validated by hand. His contributions were merged by maintainers at NVIDIA, Qiskit and Qibolab, and he topped the Wincent DragonByte 2026 qualification round, first of 492. The habit behind all of it: measure, verify against a reference, and say plainly what was not shown.

selected_work

Experiments that left the lab.

Inference engineering, AI applications, and tooling for agents. Every item links to its evidence, and says what it does not prove.

● technical preview · Sep 2026

EXP_01 · flagship

A 27B ternary LLM at up to 284 tokens/s on one RTX 5090

A language model whose weights are only −1, 0 or +1, served through an OpenAI-compatible API on a single consumer GPU. The work is the engineering between model and silicon: custom Tensor Core kernels in CUDA C++ with CUTLASS/CuTe, a faster Hadamard transform, bounded prefill memory, and tuned speculative decoding inside vLLM.

prior profile

~230 tok/s

this work, peak

284 tok/s

1.28–1.43×faster matrix kernels in microbenchmarks

2.37 GiB → 40 MiBper prefill temporary

73,089tokens in one completion, answer verified

1,585,920output values bit-identical to prior kernels

Observed on short code probes with one active request; 246–284 tok/s across restarts. The model and the original plugin are upstream work; this project is the kernel optimization, integration and validation. Implementation not published yet.

results & method ↗

verified long reasoning runs
task	generated
Constraint solving, exact probability 1/26	6,076
Circular strings, 1,038 rotation classes	27,346
All 16 weight counts, with proof	41,073
Four lengths: 272 of 272 values, with proof	73,089

tokens generated, reasoning included · answers checked against exact formulas and an independent C++ enumerator · selected successes, not an accuracy score

EXP_02

A 128B model on a 32 GB GPU

Proof of concept that runs the text portion of a 95.2 GB NVFP4 checkpoint on one RTX 5090 by keeping most weights in pinned host memory and streaming them through native C++ NIXL/UCX requests. Transfers reached 28.4 GB/s, matching a direct CUDA copy. Proof of operation at 0.38 token/s, not a serving benchmark.

NIXL / UCXvLLMC++

view repo ↗

EXP_03

Research Desk

An AI research workspace: upload documents, ask questions, and follow every citation to its passage and page. Hybrid BM25 and vector retrieval with local embeddings, streamed answers, image understanding, and saved research that reopens offline.

RAGFastAPIReact

view repo ↗

EXP_04

Live Translate

Speak in one language, hear another. A browser app that streams microphone audio and returns translated speech with a readable transcript, in 29 spoken target languages. Installable on phones, tested in Chromium and mobile WebKit.

Speech AIWebSocketsPWA

view repo ↗

EXP_05

MCP Token Relay

Agents call tools on behalf of users, and that is where credentials leak. This relay coordinates OAuth tokens across stateless MCP replicas: 64 concurrent callers produce exactly one accepted refresh, and an uncertain outcome is quarantined instead of retried. Built by several models reviewing each other under human direction.

MCPOAuth 2.0Go

view repo ↗

EXP_06

gpt2-small-czech-cs

One of the first public GPT-2 models for Czech, fine-tuned from English GPT-2 on Czech Wikipedia over a weekend on a single GTX 1080 Ti. Released in 2022 on Hugging Face under CC BY-SA 4.0 with model notes and provenance.

NLPFine-tuningOpen model

model on Hugging Face ↗

EXP_07

Multilingual GPT — 2020

A GPT-2 based conversational app with neural translation across Czech, English, Arabic, Polish and German, later grown to 11 languages with question answering and summarization. Built and deployed in 2020, two years before ChatGPT.

LLMTranslation2020

view repo ↗

EXP_08

Qualification Round 2026 — Winner

First of 492 on the final scoreboard of the Wincent DragonByte 2026 qualification round, a 48-hour algorithmic coding challenge. The same skill, breaking a hard problem down under constraints, is what kernel and inference work demands.

CompetitionAlgorithms#1 of 492

view scoreboard ↗

EXP_09

Agents for quantum computing

Where AI serves quantum: typed agent orchestration with LangGraph and Pydantic AI drives an auditable quantum benchmarking pipeline, and an MCP tool lets AI agents validate and simulate quantum circuits through one structured interface.

LangGraphPydantic AIMCP

see the quantum page →

EXP_10

WisprFlow: private dictation

Press a key, speak, and the text is typed into whatever window has focus. Speech-to-text runs on your own GPU box through an OpenAI-compatible endpoint, so audio never leaves your network. Tauri and Rust, for Linux (X11 and Wayland) and Windows.

Speech-to-textRustLocal AI

view repo ↗

$ ls ~/lab --moreGPU libraries · AI services · ports

process

From bottleneck to measured gain.

How an engagement runs, whether the goal is faster inference, a model that fits your hardware, or an AI application people can trust.

01 · profile

Measure first

Trace where time and memory actually go: kernels, transfers, prefill, retrieval. Fix a baseline before changing anything.

02 · build

Optimize or build

Custom kernels and serving changes for inference, or the retrieval, streaming and tool layers of an application. Slower variants are measured and rejected.

03 · verify

Validate

Numerical agreement against a reference, exact-answer evaluations, retrieval test sets, and failure scenarios that are made reproducible.

04 · deliver

Hand over

Pinned versions, exact commands, raw samples, and a report that separates what was measured from what was not established.

proof

Reviewed by people who ship hard things.

No slogans here. Each claim on this page was evaluated by someone else, and each one links to the record.

ranked · final scoreboard

First of 492 contestants in the Wincent DragonByte 2026 qualification round, a 48-hour algorithmic challenge. Results are final and public.

#1Wincent DragonByteQualification Round 2026

view_scoreboard ↗

code review · maintainers

GPU work merged after maintainer review at NVIDIA (CUDA Core Compute Libraries), Qibolab and Dynamiqs. The diffs, tests and review threads are public.

PROpen-source maintainersNVIDIA · Qibolab · Dynamiqs

open_the_log ↑

recommendation · IBM

Recommended on LinkedIn by an IBM manager. Benchmarks here ship with raw samples, exact commands and stated limits, so the numbers can be rerun.

IBIBM · published artifactsLinkedIn · GitHub

read_recommendation ↗

Let's run an experiment together.

Open to inference optimization, GPU kernel work, RAG and agent builds, and private on-premise AI deployments. Describe your problem, or reach me directly below.