Work below the software layer.

My products have always sat on physics I took on faith: radio links, sensor noise, orbits. In May 2026 I started working through that substrate deliberately. Primary sources, hardware on the bench, one layer at a time. I already operate the middle of this stack in the field (424 thermal-drone missions at a hundred meters over Bavarian meadows), so this page works outward from there: down to millimeters, up to orbit, and into the layer where sensing becomes a decision.

None of it was product. That was the point: a notebook of a deliberate exploration, run from May to July 2026 and then closed. The structured courses below were generated and run with tutor, a Claude Code plugin I built for exactly this kind of learning.

Millimeters · 3D Printing & CAD

How does an idea become a part that fits?

A QIDI Q2 arrived in early May, and SolidWorks alongside it. Modeling from zero, one part at a time. The printer’s main customer was the bench itself: a printer riser, an air-blower adapter, a desiccant storage box for the filament dryer, and an enclosure for an ESP32 dev board. The bench printing for the bench.

The early lesson: printing is mostly materials discipline. Filament is hygroscopic and prints badly wet, so drying became its own small experiment: a dryer, airtight desiccant boxes, trial runs across PLA, PETG, and ASA at different temperatures and humidities. Screw threads want heat-set brass inserts, not cut plastic. A soldering station covers the electronics side of the same builds.

Centimeters · UWB Ranging

How far away is my dog on our walks, in centimeters, live?

The dog is Loki, a Labrador working his way through Bavaria’s hunting-dog exams. The question cost €149 in dev boards and opened the deepest thread on this page. Two Ai-Thinker BU04 boards (Qorvo DW3000 radios) did live two-way ranging on my desk: distance readout over serial, accurate to about two centimeters, checked against a tape measure laid between them.

The boards range out of the box; understanding them was the actual project. I wrote myself a nine-phase curriculum (embedded C, RF propagation, chip-level radio work) with stop-and-verify gates on every lesson, and worked it on a bare-metal STM32: boot model, linker script, memory-mapped registers, OpenOCD debugging. No HAL, no Arduino, no generated firmware.

• Phase 0: Setup, tooling, foundations
• Phase 1: Embedded & C fundamentals
• Phase 2: RF & signal foundations
• Phase 3: DW3000 chip basics
• Phase 4: First range: SS-TWR
• Phase 5: Better range: DS-TWR + antenna-delay calibration
• Phase 6: Field-test & harden
• Phase 7: Display
• Phase 8: Direction: PDoA / AoA (stretch)
• Phase 9: Doppler primer (diversion)

The hardware delivered its first lesson early: I assumed the BU04 carried an ESP32, the documentation said STM32F103, and the whole curriculum repivoted from ESP-IDF to bare-metal STM32 within a day.

Local Spectrum · SDR Receiving

What is actually in the air around me?

A HackRF One with a PortaPack runs standalone, no laptop needed. Received and decoded: aircraft transponders (ADS-B), ship traffic (AIS), weather satellites, ham voice, broadcast, plus a lot of waterfall time learning to tell a real signal from an artifact. The waterfall is the map, and reading it is the skill. I kept analyst-style cheat sheets as I went: bands and modulation, a discovery workflow (“there’s a blip, now how do I hear it?”), decode modes, antenna choices.

The rules matter here. Receiving is broadly legal in Germany; transmitting without a licence is not, so the radio stayed in receive and I worked through the Class A amateur-licence syllabus, as much for the RF theory as for the licence itself. The active side of the spectrum stayed theoretical for the same reason: a 17-chapter electronic-warfare curriculum, from spectrum fundamentals up.

Orbit · SAR & Thermal

My drone covers a hundred hectares in a morning. What does the same problem look like from 500 kilometers up?

Thermal imaging is already my operating domain; fawn rescue lives and dies by heat contrast in the half hour around dawn. This thread worked the same physics at orbital scale. The hands-on start: I flew a photo grid with my drone, built an orthophoto in WebODM, and overlaid it on Sentinel-2 imagery in QGIS. A practice run for the workflow, and a visceral lesson in what 10-meter pixels do and don’t show: in a single frame, my own imagery at centimeter scale sitting inside a Sentinel-2 tile.

Alongside the hands-on work: NASA ARSET trainings (fundamentals of remote sensing; thermal remote sensing for urban heat islands), a nine-chapter SAR course, and the physics underneath: why thermal bands sit where they do, what the atmosphere lets through, what land-surface temperature actually measures.

Analysis · Intelligence Tradecraft

How does sensor data become a decision someone will stake something on?

Every rescue-season morning I run a small version of this loop myself: fields reported, drones tasked, finds triaged, crews dispatched. The literature reads less like theory than like scaled-up versions of problems I’ve already had. The reading program: Cynthia Grabo’s Anticipating Surprise on warning intelligence, Mark Lowenthal’s Intelligence: From Secrets to Policy, Robert Clark’s target-centric Intelligence Analysis, and Biltgen & Ryan’s Activity-Based Intelligence. Alongside the books I kept the domain itself as a linked graph (missions, functions, platforms, payloads, sensors), 140 notes deep, so every new concept had to earn a place in an explicit structure.

The thread I kept pulling was sensor fusion: how observations from different sensors, none sufficient alone, become one picture somebody can act on.

---

The notebook stays up as the record of the walk. Going down to the layer where the answer actually lives, then staying there until it makes sense, is the method rather than a detour. I’m running it again now, on a different question.