Eperi uses Privatemode to run coding agents confidentially

“Our lead developer uses Kimi K2.6 via Privatemode AI to complete tasks that had been overdue for a long time. Because our sensitive code is not exposed to third parties, we can use the coding assistant without second-guessing which tasks are safe to hand over.” - Dominik Kowol, CTO eperi

Challenge: Using AI coding assistants without exposing sensitive source code

eperi is a cybersecurity provider with customers in highly regulated sectors such as healthcare, defense, financial services, and the public sector. For CTO Dominik Kowol, this creates a clear boundary: sensitive source code must never end up with third parties. At the same time, the efficiency gains from AI coding assistants are too significant to ignore. eperi therefore looked for a way to combine both: code confidentiality and the productivity of modern AI-assisted software development.

Solution: AI coding with a verifiable confidentiality guarantee

eperi evaluated Privatemode AI, an end-to-end encrypted inference API that keeps prompts, codebase context, and model responses inaccessible to third parties. Sensitive source code therefore remains protected throughout the entire interaction, including from the vendor, Edgeless Systems, and the infrastructure operator. Unlike a data processing agreement (DPA), which is a contractual promise, this confidentiality is based on confidential computing: a hardware-rooted technical guarantee that users can independently verify. The solution integrated directly into eperi’s existing development workflow, since Privatemode AI works with common coding assistants through OpenAI- and Anthropic-compatible interfaces.

Result: Two person-days of work completed for under €5 in API costs

In an extensive testing phase across nine models from different providers, Kimi K2.6 and gpt-oss-120b delivered the strongest results through Privatemode AI. Both models produced outputs that passed all compiler and functional checks in the test. Several public models, by contrast, failed due to compiler errors or blocked legitimate domain-specific inputs, such as very large numeric values that they incorrectly classified as invalid.

The cost effect made the leverage especially clear: a task that would have taken around two person-days was completed for under €5 in API costs. eperi’s lead developers were also able to finish tasks that had long been overdue, without sensitive code ever leaving their protected environment.