Carlos Sanchez is a principal scientist at Adobe Experience Manager, the creator of the Jenkins Kubernetes plugin, and a member of the Apache Software Foundation. His work spans build automation, cloud infrastructure, and agentic applications: provisioning environments for software builds, investigating failed deployments, and assembling websites around a visitor’s intent.
From Maven conventions to Kubernetes build infrastructure
Sanchez’s open-source work includes Apache Maven, Jenkins, and Kubernetes. In 2006, he co-authored Better Builds with Maven with John Casey, Vincent Massol, Brett Porter, and Jason van Zyl. His announcement of the freely released book identified its authors as Maven core developers. In his March 7, 2012 writing about DevOps, he described Maven’s goal of automating and standardizing build steps and the practical consequence: builds no longer had to depend on the person who knew every manual step, freeing engineers to work on continuous integration, quality checks, and useful metrics.
The Jenkins Kubernetes plugin applies that approach to build infrastructure. It provisions Jenkins agents in Kubernetes pods, with pipeline-defined templates specifying the containers and tools a build needs. Teams can allocate build environments on demand rather than maintain all their capacity as a fixed collection of machines.
Audience of One: websites assembled around visitor intent
His work on Audience of One brings automation to website personalization. In his 2026 demonstration, Sanchez showed how browsing history, time spent on pages, and search queries can inform a page assembled for an individual visitor. Marketers define personas or intent groups in natural language; the system uses those groups and the visitor’s signals to choose content, recommendations, and the order of page blocks.
The team builds a retrieval corpus from the website’s own content and uses it to ground generation. Selected elements—such as hero text, product recommendations, blog feeds, navigation, and calls to action—can change while the site retains its design structure and brand constraints. This reduces the need for marketers to author thousands of variations while keeping generation tied to material the site already contains.
A coffee-shopping demonstration makes the mechanism concrete. A visitor searching for a coffee machine to use while camping receives a page combining suitable products with camping advice. A second demonstration uses AI Engineer’s site content to generate a side-by-side conference comparison for someone deciding between events. Both examples organize existing information around a decision the visitor is trying to make.
Evaluating accuracy, latency, and inference cost
For Sanchez, accuracy and latency belong in the same evaluation. His team uses Promptfoo to compare prompts across models and providers, repeating evaluations because site size, subject matter, and commerce requirements affect which model works well. He targets page generation within roughly one or two seconds and argues that a model needs enough capability for the task without imposing an unacceptable wait. Recommendations can also be generated while someone browses and prefetched before they open the page. That relaxes the immediate latency requirement, but keeping recommendations current as intent changes adds inference cost.
Agents for rollout analysis and repair
Reviewed deployment repairs: Sanchez also applies agents to deployment decisions. His AI analysis plugin for Argo Rollouts investigates canary releases and returns structured recommendations about promotion or rollback. In his 2025 work on self-healing rollouts, a failed deployment can trigger a GitHub issue for a coding agent, which proposes a repair through a pull request. After human review and merge, the change enters the rollout process again. The loop connects failure detection to a proposed fix while retaining review of the repair and scrutiny of the analysis system’s reliability.
Multi-agent investigation: In a July 2026 experiment, he divided canary investigation among log, metrics, and events analysts coordinated by a lead agent. A broken release produced reassuring resource metrics, but the log analyst found errors and opposed promotion; the coordinator accepted that objection. His case for multi-agent disagreement rests on specialists examining different signals: a healthy-looking metric should not erase evidence of failure elsewhere.
Testing sandboxes against the application lifecycle
His Kubernetes sandbox comparison takes the same care with infrastructure claims. Using Playwright and Chromium, he tested tenant routing, WebSocket connections, startup, and browser state. He distinguished preallocated warm pods from environments created on demand and examined claims about checkpoint restoration. Where a fast result reflected cached startup resources, he did not treat it as proof that browser memory had been restored. The comparison asks whether a sandbox supports the application’s actual lifecycle, including the state it must preserve, rather than judging it by startup speed alone.
Carlos Sanchez shows how Adobe’s prototype turns visitor intent into a purpose-built page by changing selected content blocks, grounding them in the existing site, and treating latency as part of model quality.
Personalize selected blocks rather than freely generating the whole website; use the existing site as the retrieval corpus so adaptations remain connected to approved content.
Treat latency as part of model quality. In the reported 15-prompt example, the chosen configuration averaged 1.1 seconds versus 4.6 seconds for the next option, but that comparison does not establish universal provider performance.
Pre-generating a “For You” page moves inference ahead of navigation, reducing visible waiting at the cost of repeated calls and potentially wasted generation.
The camping-coffee and conference-comparison demos show the central capability: reorganize several grounded blocks around a visitor’s task, rather than returning a conventional list of search results.
Behavioral personalization requires more than model engineering. Because the prototype records visited pages and time spent, a production design also needs explicit privacy and data-governance decisions.
▶ Watch ↗