I built Kidi, a C++23 inference toolkit with shared model code across CPU, Apple Metal, and WebAssembly CPU. It supports Gemma 4 E2B text generation and RTG translation, including the 500-language many-to-English model, without PyTorch or Transformers at inference time.

Models use reusable tensors and neural layers with eager execution and backend-specific operators. Gemma’s original Safetensors checkpoints load directly, including trained mixed 2/4/8-bit mobile-QAT weights, without offline checkpoint conversion.

The experimental browser demo is my first attempt at WebAssembly; it’s still slow at the moment. It uses CPU inference with SIMD and optional threads, with no WebGPU acceleration. Prompts stay on your device. The first run downloads about 2.49 GB of model files, cached for reuse. Use a current 64-bit Chromium browser with ample memory; the Wasm heap can approach its 4 GiB limit. Chat is currently text-only.

• Code and native quick start

• Experimental browser demo

• Architecture

I lead the WMT Model Compression shared task, now in its second edition, and build the open infrastructure used to evaluate complete participant systems. The WMT25 harness standardized offline Docker submissions; WMT26 adds self-contained runnable systems, common install and inference contracts, sanity checks, reproducibility recipes, and organizer-run quality, footprint, memory, and speed measurements on controlled H100 hardware.

The second edition received 41 participant-submitted variants from 13 teams. Our forthcoming findings report shows why model compression is a systems problem: reduced precision only becomes fast when the serving runtime exploits it, and efficiency claims must account for quality.

In parallel, I am building Tahoma, a C++23 runtime and research platform for language models and machine translation. Current work includes low-bit quantization, custom CUDA and ROCm kernels, paged attention, continuous batching, and multi-GPU execution.

• WMT26 shared task

• WMT26 findings report - forthcoming

• Evaluation harnesses: WMT26 | WMT25

• WMT25 findings report

• Tahoma documentation

• Quantize, Qualify, Rerank - WMT26 system paper under review

• Is 4-Bit Quantization Truly Lossless?

I created MTData to replace brittle, one-off data scripts with versioned dataset recipes. As a WMT General MT organizer from 2022 through 2026, I have maintained the official MTData setup for five consecutive editions, giving participants version-pinned, command-line-reproducible definitions of the constrained-track training data. Coverage grew from 11 WMT22 recipe IDs to 21 WMT26 recipes.

• WMT recipes: 2026 | 2025 | 2024 | 2023 | 2022

• MTData documentation and dataset search

• Many-to-English tools, data, and 500+ language model

• Quantized and integrated translation models for offline translation in Microsoft Edge.

• Built PyMarian, a Python interface to Marian’s C++ inference and evaluation engine (EMNLP 2024 system demonstration).

• Co-developed SotaStream, a streaming data pipeline for on-the-fly sampling and augmentation during model training.

• Built pigzpp, a thread-safe C++23 compression core with Python, Go, Rust, and WebAssembly bindings, plus ZIP and PNG support.

• Reworked BPE learning from repeated linear scans to heap-based updates and merged it upstream into Google SentencePiece.