2D Ising model · Metropolis Monte Carlo · simulated live
Building agentic AI systems — LLM pipelines, autonomous content platforms, encrypted-CNN vision. Physics background.
2024–25
…
A CNN classifying camera frames it can never decrypt — homomorphic encryption plus deep learning, published at SPIE. Watch it run:
stack: Python backend · React/TypeScript frontend
[1] Proc. SPIE 13606, 136060W (2025) ↗
open the full demo ↗now
…
contentgen, my content platform: one line of intent becomes a published post — LLM planning, image generation, rendering, scheduling. A real run, replayed:
[2] § 05 — the system
see the full run ↗in parallel
I’m midway through a physics master’s, and the lattice behind this page is from my thesis and first paper. Deep learning and agentic systems are the fascination I couldn’t shake — the same energy function that orders these spins trained the first neural networks, and that’s how I fell in.
[3] E. Ising (1925) · J. J. Hopfield (1982) · Phys. Rev. Research (2025)
§ 01 — Who
I build agentic AI systems end to end. My longest-running project, contentgen, turns a one-line idea into published social-media content — LLM planning, image and video generation, rendering, scheduling. In parallel I’m finishing an M.Sc. in physics at RWTH Aachen, focused on quantum technology.
I spent a year as a working student in AI Security at the DLR (German Aerospace Center), published research on the critical behaviour of the 3D Ising universality class, brought CNN inference onto homomorphically encrypted data, and picked up a certificate in quantitative modelling in finance along the way. On the side I build agentic systems — autonomous pipelines that generate, refine, and publish content without me watching.
It’s all the same interest, really: systems with many interacting parts, and models for what they’ll do together. The agentic stuff scratches the same itch — I just like building things that keep running after I close the laptop.
§ 02 — The stack
languages
ML / AI
systems & web
played with
everything here is in a repo you can read — github.com/spyder73
§ 03 — The system
My longest-running build: a fleet of microservices that takes a one-line idea and, with no further input from me, turns it into published social-media content. A studio feeds a Go pipeline engine that fans out over checkpoint graphs — LLM prompting, image, video and audio generation via provider APIs — then renders finished clips and schedules them onto platforms.
It has since grown into a multi-user platform: logins, credits metered against real model pricing, node-graph generator profiles, and a character lab that trains LoRAs on cloud GPUs. Under the hood it uses the patterns agent work actually needs — structured tool calls, LLM-as-judge validation, per-checkpoint cost metering. It’s a private tool by design — I’m happy to walk through a live demo. The core stays closed, but the public pieces are on GitHub.
studio
React · TypeScript
Pipelines are edited visually or generated by an LLM from plain language; progress streams back live.
pipeline engine
Go · Echo
Executes checkpoint graphs, orchestrates every service below, streams events back to the studio.
providers
external APIs
OpenRouter · Runware · Vertex AI · Suno — chat, image, video and audio models, selected per checkpoint.
the engine conducts these services
model registry
privateFastAPI
One source of truth: every model, its request schema, and the credit pricing the platform bills against.
store service
public ↗Python · Postgres
Persistence for pipelines, prompts, media, clips and series.
render service
privateRemotion · headless Chrome
Clip metadata + assets in, 1080×1920 reels and carousel stills out.
scheduler / uploader
public ↗Python
Accounts, scheduling, and the actual posting to social platforms.
GPU bridge
privateFastAPI · Modal
A character lab that trains LoRAs on cloud A100s (SDXL, Z-Image) and serves warm inference.
the draft — one quantum circuit
after approval — the production line
send to production
this is the step that costs
generate final media
image · video models
remotion render
assembles the clip
schedule / publish
to social media · optional
the draft — one quantum circuit
↓ idea payload + reference media
after approval — the production line
send to production
this is the step that costs
generate final media
image · video models
remotion render
assembles the clip
schedule / publish
to social media · optional
¢ — this call is metered against credits · $$ — money leaves the wallet here
Every checkpoint chooses its own model from the registry — Gemini for planning, GPT Image for stills, Kling for motion — and any of them can be swapped per checkpoint without touching the pipeline itself. Pipelines are saved as templates and run from a single line of input; demo 04 and demo 05 show two of them. If you want the full shape of the platform, the deep dive ↗ traces one idea end to end.
§ 04 — Built things
scroll →
P.01 · TypeScript · Runway API · Remotion
My Runway API hackathon submission. Describe your life story and an LLM structures it into scenes; Runway generates the imagery, animates it, and voices the narration; Remotion cuts the film. There’s even a live AI director you can talk to instead of typing.
P.02 · PyTorch
My own implementation of the GPT-2 decoder, written from “Attention Is All You Need” and trained on Tiny Shakespeare — the same attention formula you scrolled past in the bridge.
P.03 · Monte Carlo · PDE
Option pricing by Monte Carlo simulation of the underlying — then again by transforming the Black–Scholes equation into a diffusion equation and solving it numerically, as a check that both give the same price.
P.04 · C++ · Monte Carlo
The simulation behind my bachelor’s thesis and Paper I: large-scale C++ Monte Carlo of the 3D Blume–Capel model at its surface. Its 2D cousin runs in the lab on this site.
P.05 · Go
Paste an Instagram URL and an LLM judges the claim: the post is scraped, the core claim extracted, and a judge prompt returns a rating with its reasoning. A small end-to-end exercise in LLM-as-judge evaluation, in Go.
P.06 · ETHGlobal Brussels ’24
Hackathon dapp at ETHGlobal Brussels 2024, built on Scaffold-ETH 2 — NextJS, Wagmi, Hardhat, TypeScript. The weekend that got me into web development.
P.07 · PyTorch
My own implementation of He et al., 2015, trained on CIFAR-100 — ResNet32 lands 67.9% test accuracy. The residual block is the whole trick; the demo walks through why depth needs the skip.
§ 05 — Research
What happens to a critical system when you give it an edge? We studied the 3D Ising universality class in the presence of a surface — realizing its ordinary, special, and normal boundary classes — and combined high-precision Monte Carlo simulations with conformal field theory to extract boundary operator product expansion coefficients.
It’s the same lattice you watched order itself at the top of this page, taken to three dimensions and pushed to publication-grade precision.
the boundary operator product expansion
The goal was a camera that raises an alarm without anyone ever watching the footage. So we run convolutional neural network inference directly on homomorphically encrypted data: the model classifies video it can never decrypt, and plaintext never leaves the camera.
This is the kind of security-critical AI I spent my DLR year on: deep learning under cryptographic and deployment constraints.
inference without decryption
§ 06 — The map
Everything I do hangs off the same graph. Grab a node and pull — the A marks are university courses I finished with the top grade.
drag nodes · hover to trace couplings
§ 07 — The thesis
BACHELOR’S THESIS · THEORETICAL PHYSICS · GRADED 1.0
Field theory coefficients for Ising surface criticality
open PDF ↗
My bachelor’s thesis is a large-scale C++ Monte Carlo study of the three-dimensional Blume–Capel model — a spin-1 cousin of the Ising model — at its surface. It took millions of lattice updates, finite-size scaling, and more careful error analysis than I care to remember.
It didn’t stay a thesis: the work grew into my first-author publication in Physical Review Research, extracting boundary OPE coefficients of the 3D Ising universality class (Paper I above).
§ 08 — Credentials
PUBLICATION
Phys. Rev. Research — first author
Boundary OPE coefficients of the 3D Ising universality class
open ↗
CURRICULUM VITAE
CV
The one-page version — education, work, publications
open ↗
CERTIFICATE
Quantitative Modelling in Finance
Risk management · financial instruments & pricing · numerical methods
open ↗
DEGREE
B.Sc. Physics — RWTH Aachen
Thesis in theoretical statistical physics, graded 1.0
open ↗
§ 09 — Off the clock
§ 10 — Contact
I’m most interested in agentic engineering, research engineering, and AI security. Physics is where I trained; systems that ship are what I work on.
