There's a gap between what ML research claims and what production systems actually do. Most writing about AI in finance lives on one side or the other. This is an attempt to sit in the middle.
Salty Syntax started as a way to think out loud. I spend a lot of time reading papers — arXiv, SSRN, conference proceedings — and a lot of time around systems that have to work under real constraints: regulatory pressure, latency requirements, adversarial conditions, messy data.
Most writing about AI either lives in academia (clean benchmarks, synthetic datasets, controlled conditions) or in industry marketing (everything works, everything scales, everything is ready). The interesting questions sit in neither place.
What does this architecture actually look like when a regulator asks you to explain it? What does this model do when the data distribution shifts mid-year? What did the paper leave out?
That's the territory. Financial systems because that's where I work and think. ML because that's the tool. The overlap is where almost all the interesting problems are right now.
Every claim traces back to something public — a paper, a regulator filing, an institutional report. Where I'm making an argument rather than reporting a finding, I try to be clear about that distinction. Sources are cited inline and listed at the end of each piece.
The goal isn't to be comprehensive. It's to be precise about one or two things per piece that are actually worth thinking about — and honest about what I don't know.
| Area | Focus |
|---|---|
| LLMs in Finance | Where language models add genuine signal versus where they add noise. Earnings NLP, RAG on regulatory text, sentiment extraction, and the benchmark problems nobody talks about. |
| Compliance AI | AML architecture, communications surveillance, KYC automation, and explainability — the production problems that are harder than the detection problems. |
| Quant ML Systems | Time series forecasting, graph networks for transaction data, agentic pipelines, and reinforcement learning in market environments. |
| Paper Breakdowns | Recent arXiv and SSRN work read carefully, with the limitations named and the deployment implications drawn out. |
| Global Markets | Cross-jurisdictional deployment, regulatory divergence across EU, UK, US, and APAC, and what that means for AI system design. |