Zero-shot ML, secure agents and financial RAG: the day in AI

Today’s most important AI news

Monday started with little corporate fanfare but plenty of technical substance. There were no big product announcements, but something more interesting: developers and research teams showing how to do more with the models that already exist, without training anything from scratch. The common thread of the day is the democratization of capabilities that until recently required significant resources — whether for forecasting, securing industrial infrastructure or analyzing financial reports.

Zer0Fit: forecasting and classification without training a single model

A graduate student took two foundation models Google released last week — TabFM, for tabular data, and TimesFM, for time series — and packaged them into an MCP server running in Docker. The result is Zer0Fit: you connect the container to your local LLM (Open WebUI, Claude Code or Codex) and ask it for predictions, classifications or regressions on your own data without needing to build, train or tune any model.

The initial metrics are striking. On classic ML datasets like Iris and California Housing, the system reached around 94% accuracy in zero-shot mode. It’s not perfect, but it’s functional for a huge number of real use cases where a team today would spend weeks putting together a training pipeline.

The practical implication for developers and product managers is direct: tasks that used to require a dedicated ML engineer — demand projections, customer segmentation, anomaly detection in tabular data — now have a local first approximation with no training cost. Zer0Fit runs 100% on your infrastructure, which also solves the problem of sensitive data that can’t leave your organization. It’s worth keeping on the radar for quick prototypes.

AI agents controlling industrial infrastructure: promise and a clear limit

A paper published today on ArXiv proposes a framework called Neuro-Agentic Control that combines LLMs with classical control to monitor and respond to cyberattacks in operational technology (OT) environments — the kind of infrastructure that governs manufacturing plants, power grids or water systems. The problem it tackles is real: attacks on industrial OT are growing and monitoring systems based on static rules don’t scale.

The proposal is hybrid for an explicit and honest reason: LLMs hallucinate, and in a closed-loop control system that can translate into physical damage. The framework uses them for semantic reasoning and decision support, but doesn’t give them direct control over actuators. It’s the LLM as analyst, not as operator.

For industry executives and security leaders, this work marks where the reasonable limit is today: AI agents can accelerate detection and diagnosis in industrial environments, but autonomous execution on physical systems remains an unacceptable risk without additional verification layers. Anyone evaluating AI for OT should read this paper before designing their architecture.

LLMs for fundamental analysis: RAG applied to SEC filings and macro data

A team of researchers published a study on using GPT-4o with a RAG architecture to generate investment briefs from company reports, SEC filings on EDGAR and macroeconomic data such as GDP and inflation. The system preprocesses the documents, sends them via API and produces fundamental analyses that used to require hours of manual analytical work.

What sets this work apart from earlier experiments is the emphasis on real structured and semi-structured documents — not clean earnings-call transcripts, but the dense PDFs and 10-K forms any analyst knows well. The RAG pipeline handles the corpus’s heterogeneity and maintains source traceability, which is critical in a regulated context.

For product managers and entrepreneurs in fintech or wealth management, this is a sign of maturity: the architecture for automating fundamental analysis already exists, is documented and uses commercially available models. The barrier today isn’t technical but integration with existing workflows and regulatory validation. Teams that start experimenting now will have an advantage when that barrier comes down.

In a nutshell

What ties these three stories together isn’t obvious at first glance, but the pattern is clear: AI is becoming more useful not because models are bigger, but because the infrastructure around them is maturing. Zer0Fit packages Google models into something any developer can spin up in an afternoon. The industrial security framework explicitly decides where not to use an LLM. The financial analysis system solves the problem of real documents, not ideal ones. We’re at a moment where the interesting work isn’t in training new models, but in building the integration, verification and delivery layers that make existing models truly deployable. For the Latin American professionals reading this, that’s this week’s concrete opportunity.

---

Sources used: Reddit MachineLearning — Zer0Fit (https://www.reddit.com/r/MachineLearning/comments/1uue8cc/zer0fit_i_took_googles_new_tabfm_timesfm_ml/), ArXiv — Neuro-Agentic Control (https://arxiv.org/abs/2607.09076), ArXiv — RAG-Based Investor Briefs (https://arxiv.org/abs/2607.09121)