A personal engineering journal
Notes on systems, scale, and AI.
Abhishek GahlotEngineer & builderI write about infrastructure, distributed systems, and AI research. Currently training LLMs from scratch and building things that scale.
A little about meClaude Code, Codex, OpenCode and Pi all talk differently. Claude: Anthropic's SDK. Codex: a small server you send JSON to. OpenCode: an HTTP server with an event stream. Pi: a terminal. Every product that puts them in one window has to do the same three jobs, and there are only three ways to do the first one.
Read the essayThe first version saved the task and left the agent idle. Starting with one OpenCode sandbox, then building the execution and history that let four coding agents share one app.
Read the essayWhere pretraining data really comes from: the open sets, the gaps you scrape yourself, and the shadow libraries that keep showing up in training corpora.
Read the essayThree new agent frameworks shipped in 2026. I ran the same coding task on all three with Claude Opus 4.8. DeepAgents: 51s. Eve: 74s. Mastra/AI SDK: 80s. All passed 30/30. They diverge on architecture, context management, and what happens when your agent runs for 50+ turns.
Read the essayI put Tencent's CubeSandbox on a low-end dedicated box and measured every part of the spin-up path. Warm resume ~20 ms, cold create ~72 ms, a 40-wide cold burst at p50 ~490 ms (p99 over a second), a thousand create-delete cycles in 23 seconds, and a density wall at 165 concurrent. Plus how it lines up against the managed sandboxes for coding and RL, and why I ran it on bare metal.
Read the essaySeven principles from making a production LLM agent an order of magnitude faster without touching the model, and the reusable optimization prompt they became. Nothing here is domain-specific; they work for any agent that burns tokens in a loop.
Read the essayI rented an 8×B200 and tried to run GLM-5.2 on TileRT, the runtime MiMo used to push a 1T model past 1000 tok/s. TileRT doesn't support GLM-5.2, so I reverse-engineered its IndexShare attention and weight-remapped it onto the kernel. Result: ~480 tok/s, OpenRouter-identical quality, ~5× faster than vLLM on the same GPUs, capped at 2048 tokens by the closed kernel. The apples-to-apples numbers (TileRT 480 / vLLM 96 / OpenRouter 104), why MTP made vLLM slower, and what it'd take to beat it.
Read the essayA complete, apple-to-apple comparison of the three agent frameworks that matter in 2026, full feature matrix plus a real benchmark on live Daytona sandboxes with gpt-5. All three can code; they diverge on durability, cost, and architecture. Includes a pick-by-use-case guide for analytics, coding, multi-agent, RAG, voice, and enterprise.
Read the essayA technical note on Blackwood Diffusion: a masked-denoising language model with a 49,152-token agentic tokenizer, staged context growth, and revision-native generation.
Read the essayHow I designed a custom 128K Byte-BPE tokenizer for DarkForest, and the data curation strategy that turned a noisy corpus into something a model can learn from.
Read the essayNo fine-tuning. No LoRA. No base model. Training a 7 billion parameter language model from random weights, on rented H100s. General-purpose first, then specialized for code through SFT and RL on real agent trajectories.
Read the essayHow to build a chat system where multiple specialized AI agents handle different types of requests instead of one agent trying to do everything.
Read the essayWhen your AI operation takes 30+ seconds, you cant just show a spinner. Here is how to use Server-Sent Events to stream progress in real time.
Read the essayHow I added Claude Code-style team coordination to Codex — one lead agent spawning and coordinating multiple teammates in parallel using a shared filesystem task board.
Read the essayHow I built a reusable chat component library with React 19 and Vite that works across multiple projects.
Read the essay