Enforcing Compliance Policies on AI Coding Assistants with Randall Potter AWS

Presented guidance on enforcing compliance policies at inference time for AI coding assistants. The solution uses a three-layer defense architecture: OpenAI's gpt-oss-safeguard-120b on Amazon Bedrock for semantic intent classification against compliance policy documents, Amazon Bedrock Guardrails for deterministic content filtering (PII, denied topics, harmful content), and locked-down system prompts deployed on Amazon Bedrock AgentCore. The result is a Claude Code-compatible API endpoint with governance baked into the deployment image — developers use their preferred IDE extension while compliance enforcement operates transparently. Demonstrated how structural controls (IAM, guardrails) alone cannot evaluate semantic intent, and how layered policy-at-inference classification closes that gap.

Key Technologies: Amazon Bedrock, Amazon Bedrock AgentCore, Amazon Bedrock Guardrails, OpenAI gpt-oss-safeguard-120b, Strands Agents, Claude Code (Anthropic Messages API)

BIO:
Randall Potter is a Senior Solutions Architect and Generative AI Subject Matter Expert at AWS focused on putting foundation models into production at enterprise scale — with compliance, security, and governance built in as first-class engineering concerns. His current work spans the Amazon Bedrock stack — AgentCore-based agentic architectures, Custom Model Import, embedding workloads, quota and capacity planning for high-throughput inference — and the surrounding patterns that make those systems durable in production: layered policy enforcement at inference time, locked-down agent control surfaces, multi-agent orchestration, document intelligence, accessibility automation, and cost and usage observability across large multi-account environments.

His recent published work covers enforcing compliance policies on AI coding assistants at inference time, combining semantic intent classification, Amazon Bedrock Guardrails for deterministic content filtering, and AgentCore-deployed locked-down system prompts into a layered governance architecture for developer-facing agents.

A lifelong builder with more than five years at AWS and 25 years of professional software engineering experience, Randall partners closely with applied AI scientists on the engineering problems that surface when foundation models meet real production constraints.