Deep dives & threads
Long-form write-ups and threads on AI, infrastructure, and the systems I build — read right here, no login, no paywall.
Deep dives
How Sure Is "Sure"? — Calibration, Uncertainty, and Knowledge in LLMs
A hands-on experiment in calibrated confidence: a 150M-parameter student, a 7B teacher's doubts, six training targets, and three kinds of new data. What transfers, what doesn't, and why.
Steering Intelligence: A Fine-Tuning Journey
A hands-on journey through CPT, SFT, LoRA, DPO, and GRPO — what each stage changes, what breaks along the way, and what the real experiments revealed.
Autosage: Server Automation with a Built-In AI Copilot
How I built Autosage — a platform to write scripts, chain them into visual workflows, and run them on real Linux and Windows servers — plus Autobot, an AI copilot that generates, investigates, and safely executes automations, and a public docs assistant grounded in RAG.
Threads
- 01Thread
From Text to Nodes: How AI Engineering Is Evolving
From prompt engineering and context engineering to harnesses, loops, and graphs — the unit of AI engineering is gradually moving from the model to the system around it.
- 02Thread
Coherent Isn't the Same as Correct: What I Learned Building a GPT From Scratch
Built a mini GPT-v2 from scratch using a modern Transformer stack: RMSNorm, GQA, SwiGLU, and RoPE, along with a BPE tokenizer implemented from scratch. Trained on 15,000 TinyStories samples with 6.8M parameters in Google Colab, then deployed the model on Render with Gradio for interactive inference.
- 03Thread
How a Transformer Actually Learns — Loss, Backprop & Going Deeper
Training a Transformer model involves optimizing millions of parameters through concepts like Multi-Head Attention, Neural Networks, backpropagation, and learning rate optimization. These allow the model to learn patterns in data, perform tasks like language understanding and generation effectively.
- 04Thread
How Transformers Actually Work — From First Principles
Transformers are the foundation of modern AI systems, enabling models to understand language and complex data by using attention mechanisms. They analyze relationships between all parts of the input at once, allowing them to capture context, meaning, and long-range dependencies efficiently.
- 05Thread
Docker: Way to Containerization
Docker is an open-source platform that allows developers to automate the deployment of applications in lightweight, portable containers. These containers package everything needed to run an application, ensuring consistent performance across different environments.