Loading activity...
Claude Trained an "AI" for Me Claude trained a small open model to do the job a hosted 70B model used to do in my news pipeline, using answers I'd been storing all along. Training was the easy part. Figuring out if the result was good enough took longer, and we got it wrong twice.
I've Been a Naughty Developer, So I'm Punishing Myself I kept slipping past my own review guards. sdd-flow now makes me stop: a design gate, delivery tiers, eval sign-offs, tracing, and reviews by a second model.
Where AI Failed on My Provisional Patent Application Post 1 said the provisional patent was filed for $65. Here are the four places the AI work failed: the drawings, the claims, the prior-art search, and a form the patent office had to ask for.
Alleviate Your Exhaustion and Improve Your Focus When Running Agents Running several agent sessions wears you down, and tired is when you stop reading what the agent hands back. That cost me a project. What helped was removing small frictions, not trying harder.
AI Vendors Don’t Lock You In. They Lock You Out. Every vendor evaluation I've sat through was built to spot lock-in. None was built to spot the reverse: an AI system that leaves with the record of how your company works.
A Notes Plugin for People Redlining Their Agent Sessions Agents finish a task and hand you three more problems at the exact moment you are trying to merge. I researched every way to annotate an agent session, found bb, and built the plugin myself in twenty minutes.
N Newsletters to 1 Digest, Built for AI Engineers I built a pipeline to do my AI-news reading for me. It summarizes, de-dupes across sources, throws out the noise, and hands back the twenty or thirty things that mattered. It's public now.
I Sat on an Idea for 7 Years. AI Helped Me File for a Patent in 2 Weeks. Seven years ago I had an idea, 3D-printed one part of it, and then got stuck for good. Here's why I finally handed it to AI, and the honest question I wanted answered.
Securing an Always-On AI Agent at Home: A Journey Through Dead Ends The honest record of securing an always-on AI agent with shell access at home: why firewalling it in place failed, how network segmentation and an ngrok tunnel worked instead, and a frank list of what it still doesn't protect…
Our AI Rollout Started With People, Not Processes AI adoption runs on two tracks — top-down process automation and individual everyday work. We started with people: one shared, self-hosted AI tool for a sub-500-person company across several EU countries (LibreChat on Azure OpenAI in an EU region,…
Give Your Coding Agent More Autonomy I run a development pipeline with permissions off, then trace every session into LangSmith. The plain case for why: an objective record of what the agent did unattended, evals that catch model regressions before they cost me, and a loop where the agent reviews…
Which AI Should You Actually Use? I Built an AI for That. Most people use a fraction of what AI tools can do. Sherpa is an open-source agent that interviews someone about their job and recommends which AI features actually fit their work.
Claude Made an 18-Minute Tutorial of My App I pointed Claude Code at our live internal product and had it film and narrate an 18-minute tutorial. The pipeline, the parts still done by hand, and how to reproduce it.
Opus 4.7 is Terse Opus 4.7's terser default style breaks down on complex projects. The fix is a custom output style that restores thorough explanations, plus evals to catch this kind of drift on future model releases.
Thin Slices Stop /sdd-flow from Overflowing My Capacity One-shotting features with /sdd-flow breaks down on complicated projects run in parallel. The fix: thinner vertical slices that keep oversight tight. New per-slice mode is live in the plugin.
From One Agent to Many Agent coordination tools now make it easy to run multiple AI agents in parallel. A look at what's available and what Trail of Bits achieved by systematizing it across an organization.
Close Call, a Voice AI Sales Coach I built an open-source voice agent for practicing sales calls against an AI customer. Runs on Pipecat + Gemini 2.5 Flash, costs $0.14 per call.
Everything is Energy, Everything in Berlin is Beautiful On a perfect afternoon in Berlin, everyone looked happy and oblivious. I see where the train is heading, and the accident is coming. But all crises present opportunity.
Spec Driven Development Fans, You'll Love sdd-flow A Claude Code skill that orchestrates spec-driven development through subagents: research, adversarial review, specification, implementation, code review, and critical review from a single command. Built Redakt with it in two days.
Meet Redakt: Practical GDPR Compliance for AI Teams Redakt is an open-source PII anonymizer built on Microsoft Presidio. It sits between employees and AI tools, replacing personal data with placeholders before prompts leave your network. No data stored, no DPA needed, fully self-hosted.
