Havuz

Offline multi-speaker meeting-speech toolkit: planted-ground-truth synthetic meeting factory, VAD, speaker embeddings, diarization (AHC/spec

🔥🔥 Accelerating decision inference with exact state forking

Tiny-MoE is a lightweight Mixture-of-Experts language model built entirely from scratch in native PyTorch and trained end-to-end on Kaggle u

PyTorch-native framework for fault-tolerant distributed training of foundation models on AI clusters

Production-ready batch inference for Cohere’s Arabic/English ASR model, with optimized Silero VAD, multi-file GPU batching, subtitles, and o

Windows utility focused on cuda setup, nvidia local AI, and cuda environment for local AI workflows.

Watch a real AI model think - live 3D visualization of a transformer's actual forward pass: tokens, embeddings, attention, neurons, sampling

A complete, structured end-to-end roadmap to become a versatile Generalist AI Engineer from Python/math foundations through ML/DL, CV, NLP,

Object-storage-native KV cache for LLM inference & RL. Cross-restart, cross-conversation, cross-engine via shared S3 bucket.

A compact, readable PyTorch reference implementation of the DeepSeek-V4 architecture (Flash and Pro), in the spirit of nanoGPT. ~4,300 LOC.

My complete journey to becoming an Agentic AI Engineer through structured learning, projects, experiments, and production-ready implementati

Train JEV-style System 1 decision models on Qwen3.5 and Qwen3.8-27B: 24-slot decision head, KL distillation, calibration and serving. 在 Qwen

🎙️ Adaptive Audio Authentication Engine: Multi-layer voice biometric system with deepfake detection, speaker verification (ECAPA-TDNN), ASR

A stateful AI Intrusion Prevention System (IPS) that reassembles TCP packet fragments in sequence order and uses PyTorch deep learning model

NeuralForge is an end-to-end LLM fine-tuning & serving platform. It leverages LoRA (PyTorch/Hugging Face) for domain adaptation, cutting err

Profine automatically profiles and optimizes PyTorch training jobs on real GPUs, delivering measurable speedups and lower GPU costs before t

👻 MoEWatch: A lightweight diagnostic and audit library for Mixture-of-Experts (MoE) models in HuggingFace Transformers — the pytest for MoE

โคลนเสียงพูดภาษาไทยด้วย JaiTTS (F5-TTS) — ติดตั้งคำสั่งเดียว ใช้งานผ่านหน้าเว็บได้ทันที · Thai voice cloning made easy: one-command setup, l

One embedding space for text, image, video, audio, thermal, motion, and touch. Open weights, self-hostable.

I switched phones and found thousands of WhatsApp photos waiting for me, mostly memes and screenshots. This finds the ones worth keeping: se

A small chat model trained from scratch: tokenizer → pretraining → SFT → RL → multimodal. 0.78B params, CORE 0.248, tool calling, vLLM. Weig

Open benchmark for expert collapse and routing efficiency in sparse MoE LLMs

Composable LLM decoding on the probability simplex: library and paper code.

Compose neural architectures as typed, executable graph cards with two-way PyTorch sync, step-by-step local execution, and a constrained AI

Tiny, fast Brazilian Portuguese text-to-speech (21M params)

Restore old VHS and home videos on your own PC: QTGMC deinterlacing, AI upscaling, face and sound restoration. Free and open source, for Win

让 MinerU 在 AMD 显卡上以 vLLM + hybrid-auto-engine 满血运行！ 本项目是对官方文档中 AMD/ROCm 生态支持的有效补充。针对 MinerU 3.x 提供了一套完整的 ROCm 7.x + PyTorch 2.11.0 + vLLM 源码

AnyFlow - flow-matching video generation: training, inference and VBench evaluation for t2v and i2v models.

Build and train a mini Kimi K3 from scratch — a pure-PyTorch playground from a custom 88M model to 2T-scale LLM profiles.

五引擎语音合成：VoxCPM2 / IndexTTS 2.5 / 2.0 / OpenVoice / Step-Audio-EditX，声音克隆、声音设计、剧本配音、流式生成，中英日韩界面 | Multi-engine TTS (5 engines): voice cloning

Train decoder-only language models from scratch with PyTorch, featuring RoPE, RMSNorm, SwiGLU, and SDPA

Interactive, animated architecture maps for any Hugging Face model — no weights downloaded

Production-grade Machine Learning, Data Science & MLOps skills for AI coding agents (Codex, Claude Code, Cursor, OpenCode). One npx command

MoE expert compression toolkit — prune, merge, and forge specialist models for local inference.

Official implementation of "Unifying Distributional Training for One-Step Visual Generation" (MGFlow).

OpenAI-compatible multilingual TTS server — Chatterbox on vLLM with real-time PCM audio streaming, low time-to-first-byte (~0.7 s), voice cl