Scroll runs the final lap
3.948s, past the human line.
Trained from wall-stalls to a sub-4 lap; the pale line is my own 3.959s, beaten by 0.011s.
Glass House · read live from this visit
This is what I know about you without you ever telling me.
It knows where you are.
The connection carried an address but no city header, so this page will not guess a town.
It knows what you're on.
Reading the machine…
It would know you again.
Deriving the signature…
Read from this visit
- Address216.73.216.109
- Resolves toColumbus, OH
- Estimate~0 bits from 0 of 11 signals
Read in your browser, except the address rows; those travel with every connection you make. Shown back, stored nowhere.
Symphony · Applied AI, regulated finance
A ticket goes in.
An agentic pipeline that turns a Jira ticket into a reviewed merge request.
Routed by complexity.
Model routing by complexity: a lighter model orchestrates, a stronger model plans.
Built by autonomous agents.
Autonomous coding agents in isolated, ephemeral sessions, with a self-healing reconciler loop.
Proven, then merged.
A full audit trail across the issue tracker, the compute layer, and the database, kept in agreement.
- Read the case study~30%faster delivery cycle
yeGPT · ML fundamentals lab
A transformer, from scratch.
Tokenizer, attention, blocks, and training loop in raw PyTorch. No transformer libraries, no pretrained weights. 1.87M parameters.
Trained at home.
A char-level GPT trained on a Kanye corpus on a single RTX 4080. Small on purpose: the exercise is owning every layer, not chasing scale.
And this is what it writes.
At 1.87M params on a tiny corpus it produces recognizably Kanye-styled gibberish, and owning that ceiling honestly is the point of the exercise.
- Read the code1.87Mparams, zero transformer libs
All work.
Six projects, one line each. Every case study says where the work stops working.
- 01SymphonyApplied AI, regulated finance · 2025-2026~30%faster delivery cycle
- 02GhostlineReinforcement learning lab · 20263.948sfinal RL lap
- 03Glass HousePrivacy engineering · 202618+ bitsto identify you
- 04MedShield AIPrivacy engineering · 2025-2026Archived2025-2026
- 05yeGPTML fundamentals lab · 20261.87Mparams, zero transformer libs
- 06NocturneAgent tooling · 20263/3tasks shipped by its own loop
The person behindthe systems.
- Role
- Cloud Architect × Applied AI Engineer
- Focus
- Applied AI + Cloud Infrastructure
- Based
- Heidelberg / Mannheim, DE
I am 27, and when I do something I do it with full force, and it shows wherever I put that energy, whether it is the gym, gaming, or building AI systems. Underneath it I am calm and reflective, more likely to think something through than react to it. I have a soft spot for Dobermans, chocolate, and caffeine.
My current focus is applied AI: agentic pipelines, workflow automation, retrieval, evaluation, and the product layer around intelligent systems. Symphony, an agent that turns tickets into reviewed merge requests inside a regulated estate, is where most of that thinking currently lives.
Most of my best work happens between 30 Chrome tabs, late-night music, and the determination to make it work. I build best in the quiet hours, with headphones on and one problem in front of me. The taste shows up in the same places whether it is a system or a sound: restraint, timing, and the parts most people skip.
How I build
Currently: building symphony · running several projects in parallel
- Plan and concept
- Claude
- Pressure-test the concept
- /grill-me
- Implement
- /caveman + /ponytail
- Run in parallel
- /nocturne, my own skill