Developers knew AI tools but could not use them at work
Sending source code, internal documents, prompts, or sensitive context to external generative AI and public coding services was prohibited. While commercial organizations increasingly used agents for code exploration, tests, error tracing, and refactoring, internal developers performed the same work manually.
The missing element was an acceptable deployment model, not willingness
The requirement was explicit: source and prompts could not leave, inference could not depend on external SaaS, and identity, policy, models, execution, and evidence all had to remain inside. A local interface connected to an external provider was insufficient.
The deployment model changed without relaxing the security requirement
Patty placed the coding agent, model serving, control plane, and execution path inside the organization’s infrastructure. Inference requests were distributed across internal GPU resources so that concurrent use could become an everyday capability rather than a small demonstration.
The default changed from “no AI” to “yes, inside the controlled environment”
Patty Code became the developer interface. PCCP governed user identity, approved models, repository and tool permissions, sensitive context, inference routing, approvals, and audit evidence. Source and prompts were not sent outside the network.
Implementation
Phase 1 · On-premise AI development
Model serving was distributed across internal GPU infrastructure and Patty Code requests passed through PCCP policy. Automation began conservatively and expanded from observed records.
Phase 2 · Collaboration context for longer-running work
Developers, security reviewers, architects, and agents shared project context in Crew. AI activity became an identifiable participant rather than an invisible tool behind an individual developer.
Phase 3 · Specialized agents and GPU capacity operations
As 15–20 agents took on code analysis, testing, documentation, vulnerability triage, dependency review, release validation, and legacy analysis, Pilot governed responsibility, approvals, escalation, and resource limits.
Observed change
- 약 70명
- Technical staff gaining approved access to AIPreviously had effectively no approved access
- 5–7× 근접
- Throughput for selected AI-suitable workflowsNot organization-wide productivity; compares manual baselines for code understanding, tests, and repetitive changes
- 1시간 내외
- First useful understanding of unfamiliar codeSelected cases that previously consumed most of a day
- 15–20개
- Specialized agents in expanded operationManaged with ownership and GPU resource boundaries
The hardest work was not persuading engineers
Security needed to understand every path information could travel, and infrastructure teams had to keep model serving reliable under concurrency. The boundary between automatic execution and human approval was tuned gradually from actual usage evidence.
Much of the value came from understanding existing systems, not only generating code
Reading unfamiliar code, tracing dependencies, reconstructing old logic, preparing tests, and repetitive refactoring produced some of the largest differences. As agent counts grew, shared GPU capacity—not an API bill—became the scarce resource, making departmental resource boundaries an operating concern.
Evidence and disclosure
Customer-reported and operationally observed. The 5–7× range does not describe organization-wide productivity; it is an approximate comparison against prior manual processes for selected AI-suitable workflows.
The organization, systems, and identifying security configuration are withheld. Operational facts are reconstructed without direct customer quotations.