Thesis: Human-in-the-Loop Sanitization for Hybrid Local and Cloud LLM Agents under Data Sovereignty Constraints

Swedish public authorities face restrictions on sending sensitive data to cloud-based language models, while powerful on-premise alternatives require costly hardware. This thesis explores whether a smaller local coding agent can benefit from cloud assistance without exposing sensitive information. Outgoing requests must first be sanitized and approved by a human.

The assignment

• Build a workflow within an agentic coding harness that lets a local coding agent request cloud assistance through local sanitization and human approval.
• Compare one-way sanitization, where the agent adapts a generic answer to its codebase, with reversible sanitization, where stored mappings restore original identifiers in the response.
• Evaluate rule-based, LLM-based and combined sanitization approaches for residual sensitive information.
• Compare coding performance against a cloud-only setup and assess the benefits and risks of retaining identifier mappings and context.
• Measure the effort required for human review and whether reviewers detect information missed by the sanitizer.
• Assess implementation complexity, reuse of existing harnesses and anonymization tools, and local hardware requirements. OpenCode is a suggested harness, but alternatives may be explored.
• Make the cloud provider configurable and evaluate the solution using synthetic or open-source code with deliberately inserted sensitive identifiers, not real sensitive data. Classified information, legal compliance assessments and cryptographic approaches are outside the scope.

Who we are looking for

• Studying towards a bachelor's or master's degree in computer science or engineering, with graduation planned for June 2027.
• Fluent in both written and spoken Swedish and English.

Practical details

Based at Knightec Group's Sundsvall office. Start in January or March 2027. Apply through the company's careers page as soon as possible, no later than November 30, 2026. Questions can be sent to johanna.edstrom@knightec.se.