Niels Erik Nandal · AI systems engineer

Build the magic.
Keep the receipts.

I build ambitious AI products that stay legible underneath—model internals, deterministic backends, sharp interfaces, and the product loops between them.

inspect selected work

Open to full-time AI / software engineering roles

new in the notebook → go poke a token sampler in the lab

current
Fractal Tech · software engineering fellow
range
model internals → product interfaces
toolkit
Python · TypeScript · PostgreSQL

01 / selected systems

Proof, not promises.

Each project is a working argument: ambitious behavior belongs above explicit contracts, deterministic cores, and interfaces that show their work.

  • knowledge systems

    Nexus

    Production reading, notes, and AI workspace built around stable evidence, durable chat, and one shared resource graph

    ingest → anchor → retrieve → reason

    • source ↗ • live ↗

  • model systems

    niels-gpt

    End-to-end LLM pipeline from tokenizer to chat interface, laptop-first

    raw text → tokenizer → transformer → chat

    • source ↗

  • agent systems

    Factory Simulator

    Multi-agent LLM orchestration with deterministic factory scheduling simulation

    interpret → schedule → validate → trace

    • source ↗

  • generative interfaces

    Suno Session Lab: turning clip generation into a navigable space

    Turns music generation into a navigable space, layering LLM labeling over CLAP/embedding-driven clustering and centroid-guided regeneration

    generate → embed → cluster → branch

    • source ↗

open the full project index

02 / operating principles

How I make hard systems hold together.

  1. 01

    Magic above. Rigor below.

    Models interpret and generate; typed contracts and deterministic systems protect the truth.

  2. 02

    Reveal the machine.

    Good interfaces expose state, lineage, and tradeoffs so a human can steer with confidence.

  3. 03

    Readable is a feature.

    Clear seams, traces, and tests make speed compound instead of turning into debt.

03 / field notes

Thinking in public.

all writing
  1. 01
    Neural Networks from Zero: Embeddings and Backpropagation

    A plain-English walkthrough of a small next-note network: learned embeddings, hidden layers, logits, cross-entropy, and backpropagation.

  2. 02
    Xavier, Kaiming, and Batch Normalization Explained

    Why activation scale matters, when Xavier and Kaiming initialization apply, and what batch normalization actually normalizes.

  3. 03
    Building My First Real-Time Full-Stack App

    What a multiplayer Go app taught me about decomposing game logic, synchronizing state, and designing before styling.

04 / compare notes

Hard problem somewhere between a model and a real user?

I like that territory.

email me read the CVGitHubLinkedIn