Internship / Master Thesis: Quality Gate for AI-Assisted Delivery

Agentic development is becoming the default way software is built at ELCA, with agents writing a growing share of the code and architects reviewing more than they write. The challenge now is to demonstrate that the output is safe to ship and to identify who approved it. Clients such as Swiss banks, insurers, public administrations, and critical infrastructure operators are increasingly requesting transparency about which parts of a system were written by AI, with which model, and who signed them off. In this 6-month master thesis or internship starting in February 2027, you will build a quality gate for AI-assisted delivery to classify risk, record AI output provenance, define human sign-off moments, and check acceptance criteria per artefact type.

• Survey the state of the art (in-toto, SLSA, Sigstore, SPDX, CycloneDX, OPA/Rego, EU AI Act documentation) and build on established standards.
• Define a risk classification (assist, supervised, autonomous × regulated, production-facing) and the controls each class triggers.
• Specify a provenance model capturing what was AI-generated, with which tool and model, and who accepted it, automatically from commits, pull-request metadata and agent logs.
• Define human sign-off points and machine-checkable acceptance criteria per artefact type (code, tests, ADRs, infrastructure-as-code, documentation), expressed as a versioned YAML schema.
• Make the gate executable: CLI first, CI job second, with an agent self-check and a blocking pull-request check running in under three minutes.
• Write a signed gate record at each release, handle exceptions as reviewed pull requests with an expiry date, and evaluate the gate on a real ELCA codebase (false positives, runtime, developer friction).