archive

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

• buildThe list of tools I run is short. The list of tools I killed is the honest one.Twenty-one open-source tools went through my servers in four months. What survived is a short list. This is the whole graveyard, with a one-line verdict on each — including the ones I killed twice.→
• buildThe day I shut down two gateways that were working fineThree proxies was two too many. Deleting working software was harder than fixing broken software — and that's exactly how server rot starts.→
• buildThe gateway everyone recommends is 51,000 stars of someone else's codeLiteLLM is the default answer when you ask which AI gateway to run. I ran the numbers on it anyway — and picked the Go binary with a fraction of the stars instead.→
• buildI find out my AI stack is down from Slack, not from a clientFour open-source tools tell me whether my AI infrastructure is alive. If I had to hear it from a client instead, I'd have already lost the thing I was selling.→
• buildThree chat consoles look identical in a screenshot. Only two survived my VPS.Open WebUI, LibreChat and AnythingLLM all promise the same thing — one tidy window onto your AI. I ran all three on the same server, and the screenshots turned out to be the only place they were interchangeable.→
• noteA 49,000-star agent I rejected because of a 2GB billOpenHands is one of the most respected coding agents in open source. I read its source, ran the numbers, and walked away — not because it's bad, but because every task it runs needs a container I can't afford to leave lying around.→
• buildMy server has no screen and no logins. Here's how it gets into my accounts.A headless server can't click anything — it has no screen and no logged-in browser. I climbed four rungs trying to fix that. The first one hung. The one that works costs a few seconds of interruption every time.→
• noteI moved servers and the vector database didn't survive the listVector search is the fashionable way to make your notes searchable. I ran one anyway. Then a move between servers made the decision for me — and nothing broke.→
• research89% of small businesses use AI. Only 30% can show you the money.Adoption is nearly universal and results are not — a 59-point gap. The deployments that paid back share four traits, and none of them is a better model.→
• buildMy agents ask a hundred dumb questions a day. I stopped paying for most of them.Search APIs charge by the query, and agents are mostly asking cheap questions. I stood up a self-hosted metasearch engine so the paid API only handles the hard lookups.→
• 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 wrongMy server kept accepting password logins after I'd locked them off. Three separate times, the tooling quietly undid my own security work. 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.→