Uses

#Hardware

• MacBook Pro 16" (M1 Pro): the main machine, and the only one that matters when a query is running.
• Samsung 49" OLED: wide enough to keep a query and its result side by side without either one being a postage stamp.
• CalDigit TS4 Thunderbolt 4 dock: one cable to the desk, which is the whole point.
• NuPhy Air75 V2 and a Logitech MX Master 4. Low-profile keys and a mouse with enough buttons to stop reaching for the trackpad.
• iPad Pro 12.9" (M4): reading and marking up documentation away from the desk.
• iPhone 16 Pro and AirPods Pro 2, which double as the "I am on a call" signal to everyone else in the room.
• Anker Prime 100 W USB-C charger, so the desk and the bag each have one, plus a UGREEN MagFlow Qi2 25 W stand for the phone while it is charging.

#The machine that runs the models

I built this to run models locally rather than only through an API. The card is the part that decides what will load at all: 32 GB of VRAM is the difference between a model that sits in memory and one that spills into system RAM and crawls. The rest of the machine is there to keep it fed, which is what the 96 GB and the NVMe are for.

#Desk audio & video

• Rode PodMic into a Universal Audio Volt 276 interface: a compressor on the way in means calls and recordings need no cleanup after.
• Sony ZV-E10 with a Sigma 18 mm F1.4 DC DN Contemporary: fast enough that a grey Bavarian afternoon still looks like daylight.
• Elgato Cam Link 4K to put the camera into calls as a plain webcam.

#Editor & terminal

• VS Code for Python and anything that outgrows a query editor.
• Git: every analysis that someone else will rerun belongs in version control, including the SQL.

#Data & analysis tooling

• Celonis: process discovery, conformance checking, variant analysis, and Action Flows for the automation side. Underneath that, the parts nobody sees in a screenshot: Knowledge Models for the semantic layer, data pools and transformations for getting SAP data in, and YAML for the view configuration the visual editor cannot reach.
• PQL for anything inside Celonis; SQL for everything upstream of it.
• Python: pandas for analysis, and PySpark when the event log gets too big for one machine.
• Tableau for reporting that has to reach people who will never open a notebook.

#AI in my workflow

• Claude is the daily driver: reading through unfamiliar schemas, drafting and picking apart PQL, and rubber-ducking an analysis before I commit to an approach. I'm working towards the Anthropic certification for it.
• ChatGPT and Gemini alongside it, mostly to disagree with each other. Three models given the same prompt fail in different places, and the spread between them is a better signal about whether an answer is trustworthy than any one of them sounding confident.
• Local models on the machine above, to understand what a model actually is rather than treating it as a box behind an API: quantisation, context limits, and where the trade falls between a smaller model at full precision and a larger one quantised down until it fits.
• Prompting as a testable thing. If a prompt is going into anything repeatable, it gets a small ground-truth set and a score, not a vibe check.
• Where I don't use it: anything where I cannot check the answer. A model that is confidently wrong about a process metric costs more than the hour it saved.

#Everything else

• Brave for daily browsing, Chromium kept around for testing.
• Signal and Proton Mail for anything that isn't work email, and Proton VPN on anything that isn't my own network.
• Notion for notes and documentation.