A collection of things I've designed and built — from full-stack web apps to training language models from scratch.
The AutoSage ecosystem — an AI-driven infrastructure automation platform, its terminal CLI, and its docs, each backed by RAG and agentic AI.
A production full-stack AI platform that automates cloud infrastructure management through RAG pipelines and agentic AI workflows. The Django backend orchestrates long-running jobs with Redis and Celery, is containerized on GCP, and uses Supabase as the vector store.
The command-line companion to AutoSage — bringing infrastructure automation, RAG-powered queries and agentic workflows straight to the terminal, so you can provision, inspect and manage cloud resources without leaving the shell. Actively in development since August 2026.
The documentation hub for AutoSage, built on Docusaurus and backed by a Django-powered RAG assistant over a Supabase vector database — letting users ask questions in natural language and get grounded answers sourced straight from the docs.
Building and aligning language models from the ground up — each MiniGPT version pushing further than the last.
Where it started — a GPT-style transformer trained entirely from scratch in PyTorch on the works of Shakespeare, then deployed on HuggingFace Spaces with a Gradio interface for interactive text generation.
A GPT-style transformer built from scratch in PyTorch and trained on the TinyStories dataset. It layers in modern techniques — RMSNorm, SwiGLU, RoPE, GQA, LR scheduling and a hand-rolled BPE tokenizer — to improve quality and efficiency in text generation.
A fine-tuned SmolLM-135M taken through a complete alignment pipeline — Continued Pre-Training (CPT) on WikiText, Supervised Fine-Tuning (SFT) on Alpaca, preference alignment via RLHF/DPO on Orca pairs, and RLVR with GRPO for verifiable arithmetic reasoning. Each stage is a stackable LoRA adapter, served on Streamlit.
Full-stack apps and experiments built along the way.