Agentic AI, ML & GNN, financial engineering, and the boring engineering discipline that separates production systems from screenshots.
A GNN framework for option pricing that represents each option chain as a graph, contracts as nodes, edges for same-expiry, same-strike, and prior-day relationships. Across 123 days of SPX options data, the multilayer GNN beats a Black-Scholes-Merton baseline on absolute error, does not outperform a plain feedforward network, and reveals exactly what the graph structure buys and what it doesn't.
A useful reframing. Classical pricing models and modern graph-based approaches are not as far apart as the textbooks suggest — and the gap that does exist is exactly where the next decade of quantitative finance gets interesting.
I raised my per-tenant training-sample floor by 5x. The false-positive rate stayed at exactly 48.7%. Volume was never the constraint — and there's a second failure mode no sample-count check will ever catch.
Chatbots reply. Agents plan, execute, and answer for outcomes. Most of what gets shipped as "agentic AI" is a chatbot in a different shirt — and treating them as interchangeable is why so many pilots die at the pilot stage.
Most companies aren't doing AI. They're doing the appearance of AI through drag-and-drop tools that won't survive their first real production load — and they're calling it a strategy.
The model isn't the hard part. Everything around it is. The integration layer, the data plumbing, the failure modes nobody wants to own — that's where enterprise AI actually lives or dies.
Eric B. Guevarra. AI practitioner in Manila. Engineering manager. MSc, GWCPM, PSM. Twenty-five years building intelligent systems inside enterprise and financial platforms.
My working life has spanned the unglamorous middles of enterprise and financial software: trading and banking platforms, ITSM and service management, operational intelligence, data platforms carrying real production load. The kind of systems that cannot afford to be theater. That background shows up in how I think about AI. I distrust screenshots. I want the log line, the failure mode, the number that moved when the change went live.
The current work is graph neural networks for financial systems — starting with a published working paper on option pricing, extending into equity return prediction and cross-asset structure. Alongside it, applied engineering discipline on production data platforms at scale, and the field notes on this site.
Field notes are what I write when I need to think out loud about a problem I'm actually solving. Not think pieces. Not takes. Notes from inside the work.
For work I actually do — and work I'm about to do. If you're building GNN-based systems in finance, or you want to argue about whether your AI strategy is real, or you're working on production ML infrastructure, write.