I broke production twice by updating a Cloudflare token's scopes through the public API, then learned the wrangler auth fix and a secret-scrub habit the hard way. This is exactly what bit me and how I handle tokens now.
Cursor jitter and dropped video frames on NVIDIA Linux get blamed on the compositor every time. On my GTX 1660 the real cause was the driver bouncing graphics and VRAM clocks under light load. Here is the fix that held.
SQLite on one free box is one disk failure away from gone. Here is the exact Litestream-to-R2 setup I run across every PocketBase backend in my stack, including the restore drill and the gotcha that bites first.
Re-running the same deterministic prompt in dev and CI burns quota for an answer you already got. Here is the memoization wrapper I use to replay it from disk at zero tokens, and the one place I never point it.
A few daily-cron pipelines and one deploy-on-every-push workflow ate 90 percent of my free GitHub Actions minutes before the month was out. I moved CI to a self-hosted runner on a free Oracle ARM box and pushed some deploys to Cloudflare Pages native, which bills no Actions minutes at all.
I run Coolify on a single free Oracle A1 box in Mumbai as the control plane for 15-20 production apps. This is the real install, the Git-push deploy flow, the Cloudflare Tunnel that keeps every port closed, and the API quirks nobody documents.
Oracle's ARM free tier gives you 4 OCPUs and 24GB RAM for zero rupees, and it genuinely runs real workloads. I host Coolify, several PocketBase backends, and a Cloudflare Tunnel on one A1 box. Here is what actually breaks, the capacity game nobody warns you about, and the Docker bridge gotcha that breaks tunnels silently.
PocketBase and Supabase both promise a backend in a weekend. One of them bills you per project; the other is a single binary you own forever. Here is why every backend in my multi-site setup runs PocketBase on Oracle ARM, the real migration off Supabase, and the breaking changes nobody warns you about.
Workers AI runs Llama, Mistral, and Stable Diffusion at Cloudflare's edge. I tried it for a low-latency demo. This is the setup, with the rate-limit gotcha that bit me.
Coolify is the self-hosted PaaS I use across the empire. Paired with Oracle ARM's free tier, it deploys Node, Python, and Go LLM apps at zero monthly cost. This is the install.
CrewAI is the most popular multi-agent orchestration framework. I built a real research crew with it. This is the install, the workflow, and the gotcha that ate my afternoon.
I built a FastAPI wrapper around Claude Sonnet 4.6 with SSE streaming for a client. This is the structure that worked, including the gotcha that cost me an evening.
GPT4All is the simplest local-LLM desktop app I recommend to first-time users. Three platforms, one click, models running offline. This is the install plus a real first-day workflow.
Jan is the open-source desktop chat app that runs entirely on your machine. I tried it for a week as a ChatGPT replacement. This is the install plus the workflow that worked.
Langfuse is the open-source observability layer for LLM apps. Self-hosted, it gives you traces, costs, and prompt versioning without sending your data to a SaaS. This is the install.
Llama.cpp is the minimal inference engine that runs on anything. I built it from source on Mac and Linux. This is the build that worked, with the GPU flags that mattered.
LM Studio is the friendliest local-LLM front-end I have used. Built-in model browser, one-click downloads, OpenAI-compatible API. This is the install plus the workflow I use on the MacBook.
n8n is the open-source alternative to Zapier and Make. Wired to Claude, it becomes a real workflow engine for AI-augmented ops. This is the self-hosted install I run.
I ran Gemma 4 9B locally on a 32GB Linux box for a week of normal use. This is the Ollama install, the model variants worth pulling, and the one config tweak the docs do not flag.
Qwen 2.5 7B is my preferred local coding model. The pull is 4.4GB, which on Indian fibre means a 12-20 minute commitment. This is the install plus the bandwidth-saving tricks I use.
Open WebUI is the polished web frontend for Ollama. I run it as the family-friendly AI chat at home. This is the install I use, including the no-Docker path I worked out.
Pinokio is the launcher that turns 200+ AI projects into one-click installs. I tried it for image, audio, and LLM tools. This is the install plus the use cases where it actually beats manual setup.
PocketBase is the single-binary backend I use across my SaaS empire. Wired to Claude, it gives you auth, DB, files, and AI features in one Linux process. This is the install.
I ran the same content pipeline on Prefect 3 and Airflow 2.10 for two weeks each. This is the comparison plus the pick I now run for empire LLM workflows.
Phi-3 Mini on a Pi 5 is the cheapest serious LLM setup you can build. I ran one for two weeks. This is the install, the real throughput numbers, and the use cases that work.
Supabase is the default Indian-startup BaaS choice. I run PocketBase instead across the empire. This piece explains why, with the cost math and migration trade-offs.
vLLM is the production-grade inference engine I reach for when local hardware is not enough. Hosted on RunPod at pay-per-second pricing, this is the setup I use for one-off batch jobs.
Baileys is the open-source WhatsApp library that does not need the official Business API. Wired to Claude, it runs a working AI bot on your own number. This is the install.
DeepSeek at one-fiftieth of GPT pricing is real, but a full migration is the wrong move. I run it as the cheap tier inside an eight-profile llm_router with paid models behind it on failover. Here is what I route to it, what I never route to it, and why.
Local inference setup for Gemma 4 with Ollama on Mac, Windows, and Linux. Hardware requirements, performance benchmarks, and use cases where local beats cloud.