Writing

Guides and field notes for the AI-assisted team

Practical, tool-agnostic guides on AI-assisted engineering and running real production systems with small teams.

Lattice cell towers above a city at dusk, dashed detection lines from three towers converging on a single drone in the sky
Latest

Your Cell Network Is Learning to See

The same towers that carry your calls tracked drones over a Texas stadium in July - no radar added. Cell towers as sensors: what is real, what is still in the lab, the business nobody has built, and the privacy question nobody has answered. With an interactive 3D demo.

August 2026 · 15 min read
A single NVIDIA L4 GPU card on a desk with a shelf of open model names around it: Gemma, gpt-oss, Qwen, DeepSeek and DeepHat

I Self-Hosted AI for a Week. The Model Was the Easy Part.

One rented cloud GPU, a shelf of open model families, and the discovery that "the model" is an entire engineered system: the 32B wall, the harness lesson, a VRAM sanity-check calculator, verified GCP prices, and what raw models teach about guardrails.

August 2026 · 12 min read
A towering stack of notebooks and documents beside one small reusable whiteboard: the two ways an AI model can remember a long conversation

Why Long Context Is Hard: KV Cache, MLA, and Kimi Delta Attention

Storing a long chat can cost more memory than the thinking itself - 131 GB of notes for one conversation. Part 2 of the zero-background series: the growing notebook, the compressed notes, and the whiteboard that never grows. Also in Arabic.

July 2026 · 7 min read
2.8T total parameters versus about 2 percent active: a vast field of dark sleeping expert units with one small glowing cluster doing the work in a Mixture of Experts AI model

Mixture of Experts (MoE) Explained Simply

Kimi K3 holds 2.8 trillion parameters, but only 104 billion of them do any work for a given word. A zero-background guide to experts, routers, and the two numbers that matter more than the headline. Part 1 of the series; also available in Arabic.

July 2026 · 9 min read
A rain squall crossing a microwave backhaul region at dusk: links inside the storm glow amber, one burns red - a fault hiding in the rain

The Loop Meets the Microwave Network

Everyone draws the AI closed loop on RAN and core. Here is what it actually has to reason about on microwave backhaul - rain fade vs hardware ambiguity, vendor islands, blast-radius-gated actions, and the pre-reroute QoS check no diagram shows. With an open console that runs the whole loop live.

July 2026 · 14 min read
A network operations center whose wall of screens monitors AI agents: burn rates, propose-only demotion, a kill switch and an agent dependency graph

Your AI Agents Need a NOC

Telecom spent six years getting 4 percent of operators to Level 4 autonomy. AI agent teams are speed-running the same curve without the guardrails. What thirty years of NOC discipline transfers to agent fleets, what breaks in translation, and the assurance gap no tool has closed - with an interactive Agent NOC you can operate.

July 2026 · 15 min read
Agent-fleet operations architecture: a control plane, the detect-triage-correlate-remediate-escalate loop, the four control-plane pillars, and a kill switch

Who Operates the Operators?

AI agents have started causing the production incidents they were built to resolve. A field guide to running the SRE incident loop on the agents themselves: a control plane, per-agent identity, session replay, a confidence gate, and the kill switch you build before you turn on autonomy.

July 2026 · 14 min read
One repository, one shared brain illustration

One Repository. One Shared Brain.

A step-by-step guide to giving every developer and every AI coding agent the same project memory, in any environment: one canonical AGENTS.md, a versioned .ai/ knowledge base, the operating rules that keep it trustworthy, and the point where a memory service becomes worth it.

July 2026 · 11 min read