# Gabriele Mastrapasqua

> Gabriele Mastrapasqua is a freelance software consultant and open source developer based in Milan, Italy. He helps teams design, build, and improve software across AWS and cloud architecture, backend systems, product engineering, local AI, and developer platforms.

Gabriele works directly with clients, from technical decisions and prototypes through production delivery and knowledge transfer. He is also a lifelong learner, mountain walker, dog person, and blues guitar player.

## Main pages

- [Home](https://gabrielemastrapasqua.com/): Overview of Gabriele's work, interests, and selected projects.
- [About Gabriele](https://gabrielemastrapasqua.com/about/): Background, experience, and personal interests.
- [Software consultancy](https://gabrielemastrapasqua.com/consultancy/): Consulting services, working style, and the kinds of technical problems Gabriele helps solve.
- [Open source projects](https://gabrielemastrapasqua.com/projects/): Local AI, search, OCR, inference tooling, and developer tools.
- [AWS Cloud Practitioner Exam Guide](https://gabrielemastrapasqua.com/book/): Gabriele's practical AWS fundamentals book, published by BPB Publications.

## Selected open source projects

- [qwen3-tts](https://github.com/gabriele-mastrapasqua/qwen3-tts): Pure-C, cross-platform Qwen3-TTS inference with voice cloning, streaming, quantization, Metal, and CUDA support.
- [mynah-asr](https://github.com/mynah-org/mynah-asr): Native C inference engine for speech recognition and translation, with streaming and offline ASR, word timestamps, and int8/int4 quantization.
- [mynah-tts](https://github.com/mynah-org/mynah-tts): Pure-C text-to-speech engine combining an autoregressive decoder with a discrete neural audio codec, with no Python at runtime.
- [linklore](https://github.com/gabriele-mastrapasqua/linklore): Local-first bookmarks, hybrid search, and private RAG in one Go binary and one SQLite file.
- [fastapi-ocr](https://github.com/gabriele-mastrapasqua/fastapi-ocr): FastAPI-based PDF and image OCR with switchable Tesseract and PaddleOCR engines.
- [codex-lean-team](https://github.com/gabriele-mastrapasqua/codex-lean-team): A lean, adaptive multi-agent engineering setup for Codex CLI.
- [vllm-bench](https://github.com/gabriele-mastrapasqua/vllm-bench): CLI benchmarking for OpenAI-compatible inference servers such as vLLM, Ollama, and llama.cpp.

## Writing

- [Blog](https://gabrielemastrapasqua.com/blog/): Articles about open source, local AI, cloud, and software engineering.
- [Extending Qwen3-TTS: clone voices once, reuse everywhere (pure C)](https://gabrielemastrapasqua.com/blog/qwen-tts-voice-cloning/): Clone any voice from 30 seconds of public-domain audio, save it as a portable .qvoice, and reuse it anywhere — CLI, HTTP server, streaming. Pick the storage format — bit-identical (785 MB), small-and-shareable (16 MB), or postcard-sized (4 KB).
- [From RTF 3.5 to RTF 1.26: Optimizing a Pure C TTS Engine](https://gabrielemastrapasqua.com/blog/qwen-tts-optimization-notes/): How cache alignment, SIMD intrinsics, pipeline threading, and algorithm fixes nearly tripled inference speed in a pure C text-to-speech engine
- [Adding GPU backends to a pure-C TTS engine: Metal, CUDA, and the rented-Mac trick](https://gabrielemastrapasqua.com/blog/qwen-tts-gpu-backends/): How we bolted opt-in Apple Metal and NVIDIA CUDA backends onto a pure-C Qwen3-TTS engine — resident fused pipelines, server request-batching, and measuring it all on a Mac mini rented by the hour. Plus the two "obvious" optimizations we killed with data.
- [Dotfiles 2026 — Ghostty, Zed, and the joy of a modern terminal stack](https://gabrielemastrapasqua.com/blog/dotfiles-2026/): My dotfiles are on GitHub, and Ghostty + Zed are the fastest, most beautiful tools I use daily.
- [Using SQL Server on my Mac M1](https://gabrielemastrapasqua.com/blog/using-sql-server-on-mac-arm/): Let's run a docker container for SQL Server development even on mac m1
- [Emotions on a cloned voice: a 25 MB graft, a steering vector, and mixing feelings like vectors](https://gabrielemastrapasqua.com/blog/qwen-tts-emotions/): How we got sad/happy/angry/fearful speech to work on cloned voices in a pure-C Qwen3-TTS engine — a 25 MB "graft" clone that keeps the emotion levers alive, a steering + fine-tune recipe (trained on the CC-BY Emozionalmente corpus) that keeps the language stable, and — because emotion is a vector — blended "dyad" emotions plus switching emotion mid-sentence from a single prompt.
- [Qwen-TTS: Text-to-Speech inference in pure C](https://gabrielemastrapasqua.com/blog/qwen-tts-pure-c-inference/): A pure C inference engine for Qwen3-TTS with streaming, HTTP server, voice cloning, and more — no Python needed


## Contact

- [Email Gabriele](mailto:gabriele.mastrapasqua@gmail.com): Enquiries about selected consulting engagements.
- [LinkedIn](https://www.linkedin.com/in/gabriele-mastrapasqua-76ba55110/): Professional profile and client recommendations.
- [GitHub](https://github.com/gabriele-mastrapasqua): Source code, experiments, and public projects.