Hands-on guides, protocol deep dives, and field notes on tunnels, SSH, reverse proxies, self-hosting, self-hosted LLMs, and AI - from the people building Pinggy.
ZuckOff is a free app that listens for the Bluetooth broadcasts camera glasses give off and matches them against a catalogued list of manufacturer IDs. Here is exactly how the detection works, what it can't tell you, and why a solo developer built it.
In-browser LLM inference works in 2026. WebGPU ships in every major engine, WebLLM 0.2.85 carries 163 prebuilt models, and a 4-bit Llama 3.2 1B is a 695MB download needing ~880MB of GPU memory. Working code, real numbers, and the limits.
AI workflow automation: when an off-the-shelf tool is enough and when building your own is worth it. Budget, urgency, scalability and how specific your requirements are decide the answer more than the technology does.
A technical breakdown of what HIPAA-compliant AI systems actually require: BAA scope, encryption, access control, audit logging, and real EHR/FHIR integration challenges.
TypeSafe's Jev is a System One model that returns typed decisions instead of text: 70-500ms latency, $0.042 per million input tokens with output free, and 0% structured-output errors. Here are the three jobs it actually fits, the honest benchmark picture against GPT-6 Astra and Claude, and the tasks where it is the wrong tool.
Behind CGNAT your router never gets a public IP, so port forwarding cannot work. How to confirm CGNAT from your WAN address, and how to reach a Raspberry Pi or home server anyway with an outbound SSH reverse tunnel.
A technical guide to testing webhooks on localhost in 2026 - provider CLIs like stripe listen and gh webhook forward, HTTPS tunnels, relay services like Hookdeck and Svix, plus the raw-body signature trap that breaks most handlers.
MinIO's GitHub repo was archived on April 25, 2026, ending years of being the default local S3 emulator. Here's the timeline of how that happened and four solid replacements: SeaweedFS, RustFS, Garage, and S3Proxy.
Which open-weight LLMs actually fit in 8GB of VRAM or RAM in 2026, with measured file sizes, KV cache math from published configs, and Ollama commands for Qwen3.5, Gemma 4, Ministral 3, Granite 4.1, Nemotron 3 Nano, and Phi-4-mini.
What actually breaks when you run OpenRouter in production: provider-to-provider benchmark gaps, silent null responses, quantization myths, and how to configure routing that survives them.
