archive

11 entries. Build logs, notes, and research from running my own AI infrastructure. Unscheduled. Ordered by weight, not by date.

• buildRunning 13 AI agents in parallel without them stepping on each otherI got ambitious and tried to run a fleet of AI agents at once. The thing that broke wasn't what I expected — and it taught me more than the setup did.→
• buildOne shared memory so my agents never forgetEvery AI tool I use starts each session with total amnesia. I got tired of being the only one who remembered anything, so I built them a shared notebook.→
• buildWhy I moved my whole stack off the tool everyone recommendsA supply chain attack, a proxy that held entire responses hostage, and 4x the memory. What switching the boring middle layer taught me.→
• noteThe 128k trap: my AI was throwing away half its memory and I didn't know whyMy agents kept summarizing conversations at 102k tokens on models that could handle 262k. The cause was invisible defaults stacked on invisible defaults.→
• noteMost of my VPS hardening checklist exists because something already went wrongCloud-init files that silently re-enable password auth, an ssh.socket unit that overrides your config, and a heredoc that ate half an API key. Everything on my checklist is a mistake I made first.→
• noteThe marketing page is not the documentationThree times I believed a tool's website. Three times the actual code said otherwise. I'm not an engineer, but I learned to check the source — and it changed how I evaluate everything.→
• noteAdmin dashboards never get a public portI split my setup into two lanes: a public lane for APIs, a private lane for everything that should never touch the internet. The rule exists because I almost broke it.→
• logHow I actually research: two engines and a citation habitPeople ask how I produce research fast. The honest answer is a two-step pipeline plus one rule I never skip.→
• researchThe infrastructure layer is where the ROI is57% of organizations have AI agents running, but the thing holding them back isn't cost or model quality — it's the plumbing. For a small business, that changes what to spend on.→
• researchI evaluated nine enterprise AI platforms so you don't have toTwo weeks, a 28-point scorecard, 100+ references — and I'm not an engineer, so I went looking for what the marketing pages don't say. The headline numbers are mostly not real.→
• researchIndia's data center buildout is a supply chain story, not a software storyRack densities tripling, up to 42 permits to open a facility, and a country at 17% of North America's capacity. The bottlenecks are cooling, power, and paperwork — not ambition.→