Margin
My digital knowledge archive
Things I find online…and what I think about them.
Browse the archive
Follow a little curiosity.
Attention Memory Visualizer
How much memory does the attention mechanism need?
A companion to my post on the evolution of attention in LLMs. Compare KV-cache memory across MHA, GQA, MQA, and MLA as context length and batch size change.
Read the postTry the visualizer
LLM Training Memory Visualizer
How do you fit LLM training across GPUs?
See how weights, gradients, and optimizer states occupy GPU memory during distributed LLM training. Compare DDP, ZeRO, CPU offload, and gradient checkpointing to understand the memory trade-offs.
Explore training memory
Words I Kept
Quotes I didn’t want to lose.
Quotes from books, articles, and elsewhere. Browse by mood and save your favorites.
Browse quotes
Focus Mode
A Pomodoro timer with a cat for company.
Work in 25-, 50-, or 90-minute sessions, with a task list and ambient sounds.
Start a session
A SOFTER CORNER OF THE INTERNET
Take a Breather
A small pause in a very loud internet.
Breathing and soothing, for those moments when everything feels like a lot.
Take a breath
MCU TIMELINE & WATCHLIST
Marvel Universe
Find your next stop in the multiverse.
An interactive MCU timeline and watchlist with phase filters, title and character search, and release or story order.
Explore the timeline
THE NEED FOR (INFERENCE) SPEED
Model Optimization
How much faster can this thing go?
Benchmarking and optimizing deep learning models with TensorRT for NVIDIA GPUs.
Explore the code
CULTURE × CODE × CURIOSITY
AI Ganpati
When a festival meets a neural network.
Face and voice recognition brought together in an interactive festival installation.
Explore the code
PATTERNS HIDING IN PLAIN SIGHT
Crowd Counting
Making sense of a sea of people.
Exploring crowd density and flow in surveillance footage with convolutional neural networks.
Explore the code
NUMBERS, WITH A LITTLE PERSPECTIVE
Stock Portfolio Manager
A clearer picture of the portfolio.
Portfolio tracking with risk and return metrics and time-series visualizations.
Explore the code
A SMALL SCRIPT FOR A REAL PROBLEM
Vaccine Slots Notifs
A useful ping, at the right time.
A COVID-era experiment that notified people when vaccination slots opened nearby.
Explore the code
AN EXPERIMENT IN KEEPING IT SIMPLE
Resume at a Glance
The short version of the long story.
A concise, living overview of my experience, research, and work.
Take a look
The Evolution of the Attention Mechanism in LLMs
How LLM attention shares, compresses, and selects context, and what each design trades for efficiency.
Everyone Agrees With Me
How recommendation systems shape what we like, believe and expect from each other.
Winning the Metric, Losing the Point
Rethinking success in a world optimized for metrics instead of meaning
Clip, Compare, or Skip: How PPO, GRPO, and DPO Align Language Models
Why your chatbot loves saying “Great question!”
The Evolution of the Attention Mechanism in LLMs
How LLM attention shares, compresses, and selects context, and what each design trades for efficiency.
Everyone Agrees With Me
How recommendation systems shape what we like, believe and expect from each other.
Winning the Metric, Losing the Point
Rethinking success in a world optimized for metrics instead of meaning
Clip, Compare, or Skip: How PPO, GRPO, and DPO Align Language Models
Why your chatbot loves saying “Great question!”