Projects

TensorLib

C · SIMD (AVX2/FMA) · CMake · OpenMP

A from-scratch, pure-C deep learning library for CPU-only training with zero external ML dependencies (~6,000 lines, 33 unit tests).

• Dynamic reverse-mode autograd engine spanning 23 differentiable operations, with stale-graph detection and broadcast-gradient reduction.
• Composable neural network module system (Linear, LayerNorm, Multi-Head Causal Attention) culminating in a full GPT-style Transformer decoder — trained a 2M-parameter model on Tiny Shakespeare end-to-end.
• AdamW/SGD optimizers with gradient clipping and versioned, transaction-safe checkpointing for save/resume training.
• Blocked AVX2+FMA SIMD matmul kernel with packed-RHS tiling, benchmarked directly against OpenBLAS.

mAIcro

Python · FastAPI · Gemini · Qdrant · Docker

Stateless, open-source AI knowledge-management service that centralizes organizational information — open-sourced and extended by community contributors.

• RAG pipeline with Google Gemini and Qdrant Cloud, using hybrid Vector + BM25 search optimized by Reciprocal Rank Fusion.
• Real-time sync via the Discord Gateway API.

GPT-2 Inference Engine

C++ · ggml · CUDA · CMake

Pure C++ inference engine for GPT-2 (GGUF format) built on ggml.

• Weight parsing, BPE tokenization, and the full Transformer architecture.
• CUDA GPU acceleration with customizable dynamic sampling (temperature, top-k).

I-JEPA From Scratch

Python · PyTorch

Educational implementation of I-JEPA (Image-based Joint-Embedding Predictive Architecture), a self-supervised vision model trained on CIFAR-10.

• Full pipeline: multi-block context/target masking, a ViT context encoder, an EMA-updated target encoder, and a latent-space predictor.
• Validated with a linear probe well above the random baseline.

Mini-Transformer

Python · PyTorch

Lightweight English → Darija translator, trained on ~16k sentence pairs, built to learn the Transformer architecture from the inside.

ScratchNet

C++

Neural network implemented from scratch in C++.