Mike German

Led the Netherlands commercial country book for a micro-modular-reactor developer, PPAs under negotiation with data-center operators, utilities, and provincial governments, and authored the nuclear environmental licensing for the UK (ONR) and Poland (PAA). I know how enterprise and government buyers actually move, and where pilots die.

Co-founded Drinkwell, scaling a nanotech water platform to 500K end-users, 400M+ liters and $15M+ revenue Raised $5M+ in DOE SBIRs at Amorphic. Won a $925K New Mexico Advanced Energy Award for Airloom’s first commercial demo and structured a $1.6M Nevada tax abatement for Forge. I’ve carried the number, not just the slideware.

Before the AI work, and separate from the active closeout portfolio above: evaluated $1.2B+ in federal projects across 8 FEMA regions, 85 declared disasters, and 32 states / territories / tribal regions under 44 CFR Part 206 and the full NEPA / ESA / NHPA / CZMA stack. I know how civilian agencies actually buy and absorb new technology, and where they get stuck.

A federal/state offtaker-intelligence platform: real wind-resource fields crossed with the DoD installation base, substation proximity, slope, and land-status (BLM / tribal / state) to score buildable sites across multiple states and Canada, plus DIANA TEVV applications for wind-tunnel and cold-weather testing.

A multi-agent + GPU study of a barge-mounted turbine concept. The work coupled platform hydrodynamics to blade aerodynamics and structural fatigue, and identified a resonance-driven failure mode that a steady-state analysis would have missed. I then built a custom moving-base 2D CFD solver (JAX-style GPU integrator) to test the design response, and the study turned a “maybe” into a clear recommendation. Client results stay with the client; what transfers is the method.

I rebuilt an internal engineering-analysis portal: a catalogued analysis library, a gallery of rendered CFD/FEA simulation films, and a live 3-D wing-polar tool, rendered through a Vertex / Gemini image pipeline. Engineers browse the digital-twin work the way you’d browse a product catalog.

An open research program attacking one of biology’s true mysteries (no European-eel egg has ever been seen in the wild) the way you’d attack a physics problem: open data, first principles, simulation. The headline result (a stationary spawning front “warming in place”) is being prepared for a climate/marine journal, with real domain experts now in the loop. Every claim is built to be falsifiable and reported either way, including the hypotheses that died.

Read the project →

The flagship method: an end-to-end differentiable larval drift model in JAX, where the swimming policy is a neural network trained through a daily eddy-resolving ocean, with a quantum radical-pair magnetocompass in the loop, and calibrated against real arrival data by gradient descent. No published eel study does this. It runs as a GPU job in the cloud, on real ocean fields only.

I’ve written and run the “learn to build with Claude” path for non-technical colleagues across six organizations: one-word per-org aliases, siloed logins and memory, a “magic paste” self-setup prompt, and a 5-challenge teaching deck. The lesson I keep applying: for non-technical end users, the Desktop app + connectors beats the terminal every time.

For a physician with no technical background, I stood up a website she now edits herself with Claude Code (auto-deploys on save), live at integramedicalnyc.com. For Drinkwell’s field teams I built a retrieval system over their design corpus and three WhatsApp histories with recency/authority reranking, so the assistant gives the answer their lead engineer would. The goal is always they keep building after I leave the room.

integramedicalnyc.com, live →

The most verifiable proof I ship: more than 20 fixes merged into open-source projects I do not own, across 16 codebases from reactor physics (mcdc, pyne) to ocean modeling (Parcels) to bioinformatics (biopython, scikit-bio), with 60-plus more open. Each was found by a scanner, then reproduced, tested, and read by hand before it went up. Every one carries my name and a real repro.

See the merged PRs on GitHub →

A plain account of running a software operation without writing the code. The part demos skip: even a large model’s maximum-confidence findings were wrong 40 to 50% of the time. The only reliable filter was boring: reproduce it, run the tests, read the surrounding code. That gate, not the model’s confidence, is what put fixes upstream.

Read the write-up →