Francisco UtreraAI Product Engineer · New York

AboutPosts

I build AI platforms that run in production for real customers: the agent runtime, the tools it can reach, the evals and guardrails around it, and the operations that keep it honest.

Most recently I was a Staff Engineer at Syndio, where I built Syndi, the company's AI platform, from an empty repository to general availability for every customer in ten weeks. Syndi is an LLM harness with a knowledge base, file ingestion, a Slack surface, and gVisor-sandboxed code execution; an eval and LLM-judge system on Datadog LLM Observability; a per-org kill switch and a customer-data gate; usage and cost analytics; and the governance framing that let a conversational agent ship on top of pay-equity data. Other product teams build on its API.

Before that I led the AI stack at Embrace.ai for three years, until Syndio acquired it in 2026: agentic frameworks running ReAct loops over curated tool definitions with sandboxed code execution, a RAG system with vector search and semantic reranking, knowledge-gap recommendations, and a multi-modal document extraction pipeline.

Before industry I did research on adversarial robustness and noise-based training at UC Berkeley (3 papers — ICLR '21, ICLR '22, AISTATS '24), including first-author work showing that adversarially-trained networks transfer better to downstream tasks.

In November 2026 I joined Bridgewater Associates in New York.

Selected work

  • Syndi, Syndio's AI platform (2026) — Built from an empty repository to general availability for every customer in ten weeks, then led it: an LLM harness with a knowledge base, file ingestion, a Slack surface, sandboxed code execution, and an eval system that judges every production conversation. Shipped with a per-customer kill switch and data gate before it shipped features, which is what let a conversational agent launch on top of pay-equity data. Other product teams build on its API.
  • The AI layer at Embrace.ai (2023–2026) — Owned it end to end for three years: a provider-agnostic LLM SDK with multi-model routing and ReAct agent loops, retrieval over 3.5M+ documents with vector search and reranking, and an agentic tool framework for Zendesk, Jira, Confluence, and Salesforce. Served 100K+ conversations a month until Syndio acquired the company.
  • Research at UC Berkeley — First-author work showing that adversarially-trained networks transfer better to downstream tasks (ICLR 2021), and follow-on work on noise-based training for robustness (ICLR 2022, AISTATS 2024).
    © Francisco Utrera.RSS