Working with AI
Two threads. AI is my pair programmer and orchestrator — it helped me build a privacy-first journal app end to end and run a ten-round penetration test against it. And where it earns its place in a product, I build it in: opt-in, on-device by default, cloud only when a user explicitly chooses it. Either way the rule is the same — ship it, then verify it actually does what it claims.
Method
It answers questions I haven’t formed yet — diagnosing a deadlock in one pass, then explaining why it deadlocks so the mental model sticks.
I drove a ten-round penetration test from one terminal: the AI read the whole codebase, wrote exploit scripts, ran real tools across a Kali/Windows/Ubuntu lab, and fixed findings in-session.
Several investigations run in parallel — one on the network protocol, one on how keys are wiped from memory, one on access control — then the findings get pulled back together.
A fix is new, untested code. I attack my own fixes round after round until an independent hunt comes back clean — that’s how I caught an “encrypted at rest” feature that had never actually engaged.
Force multiplier
AI-assisted development is how I actually ship — here's the work that documents it.
Ring 3, a syscall gate, and the two bits that turn 'the kernel is the only program' into 'the kernel runs code it doesn't trust.' The first time in this project that the hardware, not my kernel's goodwill, holds code back.
ReadSubdirectories, cd/pwd/ls/cat, and the moment the core arc — boot, memory, tasks, shell, files — is complete. Along the way I learned the hard way that a termination proof is not a safety proof: my validator provably terminated and still hid two denial-of-service bugs.
ReadWe learned the hard way what an engagement-optimized feed does to a mind. A model trained to agree with you is a different version of the same problem — and for some people it has already turned dangerous. How to use AI without letting it become a yes-man for your worst spirals.
ReadI'm optimistic about AI and use it every day — which is exactly why I wanted to write the best possible case against it. Not strawmen: the arguments that actually keep me up, argued as well as I can argue them.
ReadThe scary 'cognitive debt' study is real but softer than the coverage — here's the honest read, plus a personal counterexample from jury duty.
ReadThe bespoke tooling that made a ten-round self-pentest possible, the attacks that failed, the full round-by-round scoreboard, and the lessons that outlive the app. Part 4 of a four-part series.
ReadNot the hype version — the actual tools, habits, and guardrails I use to build, write, and think with AI every day, and the specific ways I keep it sharpening my judgment instead of replacing it.
ReadA model people relied on got switched off overnight by an order none of its users had a say in — and why that's a reason to hedge, not to stay away.
ReadAI-written text has tells, so I keep a blocklist of the worst clichés and reject any auto-post that contains one before it can leave in my name.
ReadThe eighth round of a self-pentest found MoodHaven's flagship 'encrypted at rest' feature had never engaged on any build — then a custom attack tool and a fix-the-fixes round closed the campaign. Part 3 of a four-part series.
ReadThe LiteLLM scanner attacked the gateway. These two garak probes attack the layer above it — getting the model itself to hand you a shell command or a Mongo operator, on the bet that something downstream will run it. Here's how the probes work, and why the detectors are the hard part.
ReadI don't like posting to social media, so I built a pipeline that publishes my blog and promotes it across LinkedIn, Bluesky, and X — on a schedule, in my voice, with a linter that won't let it sound like a robot. Here's the whole system, including the two automated feeds I had to stop from colliding.
ReadI put the 'Ask Me' bot on this site through a real security pass — prompt injection, jailbreaks, input fuzzing, and an automated LLM scanner from a Kali box. Here's what held, what surprised me, and the one latent bug I found.
ReadThe confirmed vulnerabilities from a ten-round self-pentest of MoodHaven Journal: a readable database, silently lost edits, keys leaking over the LAN — and the critical bugs my own fixes introduced. Part 2 of a four-part series.
ReadA small RAG-style assistant that answers questions about me, built with no vector database, no framework, and a model bill measured in cents — here's the whole architecture, and the guardrails that keep a public endpoint from spending my money.
ReadDoes AI make us dumber, or sharper? The honest, fact-checked version — what the research really says about AI and your brain, its real environmental cost, the quiet ways it flatters you, and how to use it well.
ReadA browser remake of the 1979 Atari Lunar Lander that turned into real orbital physics, historic Apollo missions, and an AI Theater where you watch a neural net learn to land in real time. Built with AI, for fun.
ReadI'm a generalist, not a specialist — and for the first time that feels like an advantage instead of an apology. How AI changed what one IT person can actually do in a day.
ReadHow I went from a templated Wix site to an owned, fast, markdown-native stack — thought to design to a live cutover — in about a day, with AI as the pair programmer and me making the calls.
ReadWhy I ran a ten-round penetration test against my own privacy-first journaling app — the attack lab, the AI-orchestrated workflow, and the rules of engagement. Part 1 of a four-part series.
ReadBuilding MoodHaven Journal from scratch as a non-developer — zero-knowledge encryption, a Wear OS companion, peer sync, and what it actually takes to learn while you ship.
ReadIn the product
Opt-in, on-device first, cloud only on explicit consent.
Attack-surface monitoring for self-hosted and AI infra. AI shows up twice and only ever advises: it drafts a CVE detection from an advisory URL — which I review and sign before it ships — and it triages each new finding off the request path, in a durable Arq + Redis queue that no-ops cleanly without a key. A human and a signature always sit between the model and a customer host.
A consent-tiered dictation and reflection pipeline: local filler-word cleanup always on, optional on-device polish via Ollama, and OpenAI only with explicit BYOK consent. On-device sentiment detection, and Oura health context sent as qualitative descriptors only — raw biometrics never leave the device. Every AI feature is opt-in.
A browser-based game exploring game design and AI-driven mechanics, built with TypeScript and PixiJS.
Newsletter
New essays and short notes — privacy-first software, AI, security, and the occasional rambling from the trail. Every other week, no filler.