Melanie Warrick is a software engineer and co-founder of Fight Health Insurance with Holden Karau. Her career spans production machine learning, developer education, healthcare software, and durable agent systems. Across that work, she addresses what happens around a model: how developers operate it under load, how people review its output, and how an application survives failures or long waits for a decision.
From production machine learning to developer education
Warrick began in business consulting and enterprise technology implementations before moving into production machine learning at Change.org. She then became a founding engineer on Deeplearning4j, helping build deep-learning software for the Java ecosystem. At Skymind, her contributions included Deeplearning4j and its numerical-computing companion, ND4J. Her engineering and consulting work brought her into the practical work of making machine-learning models usable in products and leading teams that built them.
She subsequently joined Google as a senior developer advocate focused on AI and Google Cloud. In an extended conversation about AI and machine learning, she distinguished narrow, task-specific intelligence from expectations of humanlike machines and explained how transfer learning lets developers adapt an existing model to a new problem. Those explanations connected her production experience to questions developers faced when choosing and applying models.
In 2018, Warrick co-hosted the Google Cloud Platform Podcast with Mark Mandel, contributing to nearly 60 episodes covering machine learning, developer tools, infrastructure, and customer applications. She deliberately sought diverse voices and perspectives for the program. Her subsequent engineering career included Tailscale and Virta Health. At the AI Engineer World's Fair 2026, she presented her durable-agent work in a developer relations role at Temporal.
Making healthcare appeals easier to pursue
Warrick co-founded Fight Health Insurance with Karau to help people prepare appeals against insurance denials. Helping loved ones navigate healthcare challenges gave her a personal reason to work on the problem: a denial can impose another demanding administrative task when someone already has little capacity to handle it.
The free service turns denial information into appeal drafts that people can edit and submit. Its AI system combines fine-tuned medical language models with searches of PubMed for relevant research. Training examples are generated from public state appeal decisions, and low-rank adaptation adjusts a limited portion of a model's parameters. The resulting letter requires human review, including checks of its citations and the particulars of the patient's situation. The team builds the drafting assistance; the person pursuing the appeal retains responsibility for reviewing and submitting it.
Operating models and agents beyond the happy path
Warrick's fine-tuning guide follows model adaptation from the decision about whether it is needed through hardware constraints, quantization, training tradeoffs, evaluation, and deployment. Her durable-agent work develops a related concern: producing an answer is one step in an application that must also manage resources, preserve progress, and give people effective control.
Production behavior under constraints: A model tested on individual requests must contend with concurrent users, finite GPU memory, changing prompts, and overload once deployed. Warrick's production deployment guidance connects independent versioning of weights, prompts, and inference settings to practical recovery: a faulty prompt should be reversible without replacing the entire model service. Readiness checks must establish that the server can generate, while capacity tests reveal when accepting more work becomes harmful.
Human approval as asynchronous work: In her multi-agent systems work, Warrick treats a person's response as a decision that may arrive hours or days later. The pending request must survive a worker restart. Her “The Human Is an Async API” talk explains how a wait condition lets a workflow pause and a signal lets a person's response reach it later. She also distinguishes a person interrupting an agent from an agent requesting approval: an operator's stop instruction should reach workflow control directly rather than depend on a model interpreting it. Choosing when to request approval requires weighing the cost of a wrong decision against the fatigue caused by too many alerts.
Durability beneath agent frameworks: Her Ziggy's ice-cream delivery demonstration separates agent reasoning from execution history. Google ADK and LangGraph handle different portions of an order while Temporal preserves completed steps and pending decisions. Killing a worker during an approval wait demonstrates why that separation matters: the application can recover its progress when execution resumes. Recovery reuses recorded results, but an interrupted model call may need to run again, so tools with side effects still need safeguards against duplicate actions.
Warrick also hosts Vibe Check, exploring AI tools, agent development, and production engineering. Her teaching draws on the same practical questions that shaped her engineering work: whether a system can operate under real constraints, whether people can intervene effectively, and whether useful work survives an interruption.
Melanie Warrick’s ice cream delivery demo shows how Temporal lets an agent wait for a human without stopping unrelated work—and recover an approval even when its worker goes offline.
A human-dependent decision should pause the affected work while unrelated orders continue.
In the worker-outage demo, the separate event store queues an approval, and replay reconstructs the workflow’s state so the order can continue after recovery.
▶ Watch ↗