Spegel Explained — A Visual Architecture Guide
Visual architecture guide to Spegel, a stateless cluster-local OCI registry mirror for Kubernetes using P2P networking.
Tremo Ahmed
Notes from research engineering
AI agents, inference & ML systems
I'm Tremo. I work on AI agents and computer use at Meta Superintelligence Labs. Here, I write about how ML systems work, from serving LLMs to the infrastructure behind large-scale training.
Start here
Three guides to the systems behind the models.
- LLM inference
How to Efficiently Serve an LLM
Batching, KV cache, quantization, and the trade-offs in serving a model. - Training infrastructure
What Infrastructure Does It Take to Train a 405B Model?
Networks, storage, compute, and recovery at the scale of Llama 3.1 405B. - Distributed systems
Spegel Explained — A Visual Architecture Guide
A visual guide to sharing container images across a Kubernetes cluster.
All writing
Subscribe via RSSHow to Efficiently Serve an LLM
How batching, PagedAttention, quantization, speculative decoding, and KV cache management fit together in an LLM serving system.
What Infrastructure Does It Take to Train a 405B Model?
A look at the networks, storage, compute, and fault tolerance behind large-scale training, using Llama 3.1 405B as the reference.
The Tech Behind TikTok's Addictive Recommendation System
A look at ByteDance's published Monolith recommendation architecture: online training, streaming features, and collisionless embedding tables.
Do You Really Need a Vector Database?
Exploring alternatives to vector databases for LLM applications, including FAISS, PostgreSQL pgvector, and Elasticsearch.
Starting a Blog on GitHub Pages
Step-by-step guide to setting up a personal blog using Jekyll and GitHub Pages with custom domain configuration.