Blogs
Musings on backends, data, and whatever else is on my mind.
- NOV 12026
Amazon Customer Review Sentiment Analysis
A plain-English story of how I taught a computer to read Amazon phone reviews, tell happy customers from unhappy ones, and show sellers what people complain about most.
- OCT 292026
Saving 100TB of RAM with Math. Rebuilding Cloudflare’s Hash Ring from Scratch in Rust
Rebuilding the story of how Cloudflare cut about 100TB of RAM from its load balancer, step by step in Rust: checking every formula against real numbers and measuring what happens when the ring changes.
- OCT 232026
This Coding AI Redesigns Its Own Workshop, Then Trains Inside It
A simple breakdown of the paper “Self-Evolving Coding Agents via Harness-Policy Optimization”: a coding model that rewrites its own harness, trains inside it, and gains more than either fix alone.
- OCT 182026
Beyond the GPU: Solving the Memory Wall in LLM Inference
A simple walk through the paper “Reimagining LLM Inference Infrastructure with Memory-Centric KV Cache Servers”, and why memory, not math, is what limits most LLM inference today.
- OCT 92026
Building an Extra Security Guard Layer for AI Agents
A practical guide to making AI agents safer, be it for personal use or for industrial and corporate use.
- SEP 252026
What Is Jev? Is It a New Model or Just a Buzzword?
A look inside TypeSafe AI’s System One Model, and why it might be the missing link for real automation.
- SEP 192026
Do LLMs Really Understand Math? A Knowledge Space Theory Perspective
An LLM can solve the hard problem and fail the arithmetic underneath it, and accuracy still scores that as a win. An ETH Zurich paper uses Knowledge Space Theory to test whether a model's right answers respect their prerequisites, and finds knowledge that is fragmented rather than hierarchical.
- SEP 102026
Distributed Rate Limiting
A counter in process memory stops meaning anything the moment you run a second replica: each instance enforces its own limit, so ten pods let through ten times the traffic you configured. What it costs to keep one shared count honest.
- SEP 42026
Speculative Decoding: Making LLMs 3x Faster with Zero Quality Loss
A small, fast draft model proposes the next few tokens and the large one verifies them in a single batched pass. The speedup is 2-3x, and the output distribution is provably identical to running the large model alone.
- AUG 312026
Chaos Monkey: Breaking Your Own System in Production on Purpose
A SIGCOMM paper argues the recovery code in a network controller is only ever tested by breaking the live network on purpose: what to break, how to prove the break is survivable first, and why the bug can live entirely in the milliseconds before the system settles.