← All speakers

Al Chen

Conference affiliation: Cisco

Al Chen builds the engineering workflows that help large banks put AI observability software into production. A software engineering leader at Cisco, where he was introduced at AI Engineer NYC 2026 as leading customer engineering for Splunk Agent Observability, formerly Galileo, Al works through the security, infrastructure and organizational problems that stand between a working product and an approved deployment.

On this page

His contribution includes changing how customer engineers investigate those problems. Al developed a Claude Code customer engineering workflow that draws on Galileo’s repositories and customer-specific context to help answer technical questions and provide architectural guidance on Kubernetes clusters, LLM gateways and Helm charts. In his April 2026 demonstration, the codebase becomes working context for customer support: the assistant can use the implementation alongside the customer’s circumstances when helping an engineer reason through a question. Al also described this work in a first-person account of sharing the workflow with developers.

That work sits within a demanding customer environment. Banks run different versions of a vendor’s container images, use different vulnerability scanners and impose different remediation deadlines. A package version accepted by one bank may conflict with another’s requirements. Al’s team uses an agent skill to organize vulnerability research and reporting, distinguish fixes it can make from dependencies on open-source maintainers, and explain those constraints to bank security teams.

A defining demand in his approach to banking deployments is evidence beyond a successful status check. Updating a package, completing a backup or redacting a customer’s name in an interface can each leave the underlying risk unresolved. Al makes that distinction concrete through tests of exploitability, recovered data and evaluation outputs.

What production checks need to establish

  • Test the exploit after the upgrade. A vulnerability scanner may recognize a patched package version without establishing whether the application remains exploitable. Al proposes using a Harbor agent harness containing the application’s Docker images, instructions and tests. An agent reconstructs an exploit for the older package, tries it against the updated application and writes a test for the result. Engineers gain a way to investigate the application’s behavior beyond the version number reported by a scanner.
  • Verify the recovered data. In ClickHouse deployments using backup and restore jobs, successful job logs do not establish that the data is still correct. Replication lag, a bad disk or a failed migration can leave it stale. Al recommends comparing table row counts and checking whether replicas have converged. Recovery design also determines where records live: one pattern he describes replicates identifiers and metadata across regions while keeping core trace data within its required jurisdiction.
  • Check what judge explanations reveal. An LLM evaluating a raw trace can reproduce personally identifiable information or material nonpublic information in its explanation, even when the application’s interface redacts those details. Al describes retaining evaluation scores while withholding explanations that contain sensitive information. When reviewers still need an explanation, a second judge can evaluate redacted input. That privacy measure adds another inference call, making its operating cost explicit.

Deployment speed depends on the bank

Al challenges deployment estimates that ignore a customer’s infrastructure and approval processes. Established procedures for importing container images and GPUs ready for inference can shorten the path to production. Formal security reviews, custom LLM gateways and role-based access configuration add work; missing image-approval processes or GPU procurement plans can extend it further.

His approach to compliance velocity therefore includes human coordination alongside software. When a customer raises a concern, his team breaks it into engineering, workflow and organizational tasks. Answering architectural questions with codebase context and testing a deployment’s security claims serve the same practical purpose: giving customers enough understanding and evidence to move consequential software into production.

Read the topics behind these talks

1 conference talk

Loading more talks and highlights…

References