02
mynah-org · Speech
mynah-asr
A fast native C engine for speech recognition and translation, llama.cpp-style: streaming and offline ASR, word timestamps, int8/int4 quantization, and no Python at runtime.
Open source, built in public
Projects
Things I build to make local AI, search, and developer workflows more useful. Small dependencies, practical interfaces, and software you can actually own.
01
Featured project
qwen3-tts
A lightweight, cross-platform inference engine for Qwen3-TTS, written in pure C. No Python, PyTorch, or ONNX runtime—just C, BLAS, and raw model weights.
It runs the complete text-to-speech pipeline locally, supports the 0.6B and 1.7B models, nine voices and ten languages, with voice cloning, streaming, quantization, emotion controls, and optional Metal and CUDA backends.
03
mynah-org · Speech
mynah-tts
The same idea for the other direction: a pure C text-to-speech engine pairing an autoregressive decoder with a discrete neural audio codec, running sub-realtime on CPU with optional Metal and CUDA.
04
Local-first
linklore
Bookmarks you actually own. A local-first link manager with hybrid FTS5 + semantic search and private RAG chat with citations—all in one Go binary and one SQLite file.
05
Document AI
fastapi-ocr
A FastAPI service for PDF and image OCR with switchable Tesseract and PaddleOCR engines, automatic rotation, smart text extraction, and LLM-friendly output.
06
Agent workflows
codex-lean-team
A lean, adaptive multi-agent engineering setup for Codex CLI, with focused specialists, reusable skills, safe defaults, and bounded delegation.
07
Inference tooling
vllm-bench
A lightweight CLI to stress-test any OpenAI-compatible server—including vLLM, Ollama, and llama.cpp—and measure throughput, time to first token, and behavior under load.
The speech engines live together under mynah-org . More experiments and work in progress live on my GitHub profile .