Francisco Javier Arceo

Francisco Javier Arceo maintains open infrastructure for production machine learning and agents, including Feast, OGX, and vLLM Agentic API, and serves on the Kubeflow Steering Committee. His work spans the systems around a model: supplying its inputs, managing conversation state and tool execution, and enforcing access to enterprise data. His career includes financial-risk modeling, machine-learning infrastructure, and forward deployed engineering at Baseten.

From financial models to production infrastructure

Francisco began his career in 2012 building statistical models for marketing and insurance. His early work included AIG and Commonwealth Bank of Australia, where he helped launch TymeDigital. In 2017, he joined Goldman Sachs as an early employee of Marcus, working on credit risk, data engineering, and deploying statistical models. He subsequently built Unidos, a financial-wellness mobile application for the Latino community.

At Fast, an assignment to build risk machine learning expanded into data infrastructure, machine-learning infrastructure, a recommender system, and leadership of two teams. After Fast collapsed, Francisco joined Affirm through an acquihire. There, he managed a team building machine-learning and data infrastructure, including its feature-store infrastructure.

His 2022 production-ML essay explains why infrastructure became central to his work: a promising model could require extraordinary engineering effort before customers could use it. He increasingly worked on that deployment bottleneck himself.

His recommendations follow the ways production systems fail. Test incoming data schemas and feature transformations: an upstream type change can break a downstream model whose existence the data producer does not even know about. Separate feature engineering, numeric serialization, model execution, and business logic so engineers can isolate failures. Precompute features where possible, and introduce streaming complexity when the application requires it. Continue monitoring after deployment; Francisco observed eventual degradation in every applied industry model he had built.

Choosing open infrastructure

Francisco became a Feast maintainer while working at Affirm. In April 2024, he left fintech for Red Hat OpenShift AI. He wanted AI development to be less concentrated among a few organizations and believed structured-data infrastructure deserved more attention amid enthusiasm for language models.

Feast addresses a problem familiar from that career: keeping model inputs available across development and production. Its offline store supplies historical features for training; its online store supplies features for low-latency predictions. Point-in-time retrieval helps prevent training datasets from including information that would not have existed when a prediction was made. Francisco joined the existing project as a maintainer, contributing to a shared infrastructure effort.

His Kubeflow agenda focuses on adoption as well as capability. He advocates simpler workflows, thoughtful defaults, better documentation, and useful subsets of the platform for software engineers incorporating AI into products. He also proposed making retrieval-augmented generation a first-class priority. These choices address the work required to adopt and operate Kubeflow, rather than treating additional features as sufficient progress.

Agent execution and enterprise permissions

Francisco’s later work at Red Hat included open agent infrastructure. He coauthored the OGX paper with ten collaborators, describing a vendor-neutral application server with interchangeable inference, vector-storage, and safety providers. OGX, formerly Llama Stack, lets applications use familiar APIs while retaining choices about the models and infrastructure underneath them.

His maintenance and research work addresses several distinct problems in those applications:

• Agent runtimes beyond inference. As one of the maintainers of vLLM Agentic API, Francisco helps develop the application layer that connects coding clients such as Codex and Claude Code to vLLM and open models. The Rust server manages conversation state and multi-turn execution. Its tool-ownership model distinguishes tools executed by the gateway from tools returned to the client or handled by the inference provider. A gateway can run web search while a coding client retains its local editing tools.
• Authorization before context construction. In research coauthored with Varsha Prasad Narsing, Francisco examines a mismatch in enterprise retrieval: the most relevant document may belong to someone the caller has no permission to read. Their architecture attaches ownership metadata during ingestion, checks access during retrieval, and centralizes security-sensitive orchestration on the server. This places enforcement around retrieval, tool execution, and conversation state before unauthorized information can enter the model’s context.
• Retrieval that remains useful after access checks. The same research distinguishes filtering unauthorized results after search from applying access restrictions within the search itself. Post-search filtering can prevent disclosure yet leave too few useful results when other tenants’ documents dominate the rankings. The authors recommend backends that apply restrictions during retrieval. Their synthetic evaluation reports that attribute-based access-control gating eliminated cross-tenant leakage with negligible overhead; they also identify deployments where this architecture is unnecessary, including public-data and single-tenant applications.

Across these projects, Francisco’s contribution is to the infrastructure that lets other engineers put models to work: feature delivery for predictions, execution state for agents, and permission checks for shared enterprise systems.