About

I’m an engineer who spent the last two years building and running production systems where algorithmic trading meets machine-learning infrastructure — the messy seam where a model’s output has to survive contact with real money and real latency.

Most of that work lived in two worlds at once:

• Trading systems — systematic strategies, backtesting pipelines that don’t lie to you, execution and risk plumbing, and the unglamorous work of keeping a live book from quietly bleeding.
• LLM & GPU infrastructure — self-hosted inference stacks, multi-GPU serving, and the data pipelines that feed them, run on a budget rather than a hyperscaler’s credit card.

nullalpha is where I write those two down together — field notes from production, with the failures left in. The name is half joke, half thesis: null for the honest starting point (most edges are noise until proven otherwise), alpha for the thing we’re actually chasing.

What you’ll find here

• Trading — strategy design, backtesting pitfalls, regime detection, execution.
• AI Engineering — vLLM, quantization, KV-cache sizing, self-hosted inference, real production postmortems.
• Crossover — using LLMs inside a real quant workflow: what actually adds signal, and what’s just expensive noise.

Posts are first-person and specific. When something broke, I say how it broke.

Why anonymous?

Some of this touches live trading and infrastructure I’d rather not tie to a real name. The trade-off is deliberate: everything here should stand on whether the engineering is right, not on who wrote it. If a post is wrong, tell me — I’d rather fix it than defend it.

— Neil A. Sohn