← All selected workCASE STUDY / 01

Security & developer tooling

AgentShield.

Fast agents. Considered decisions.

A policy-as-code control plane for inspecting risky AI-agent changes before they reach production.

Public control-plane demo
STACK / PROJECT CONTEXT

TypeScript · React · Express · PostgreSQL · Prisma

Inspect the repository
SYSTEM EXPLORERSource-backed illustration
examples/vulnerable-repo

Start with the change, not a score.

An intentionally vulnerable repository supplies the demo input. No live repository or personal data is scanned by this portfolio.

Inspect a boundary. Follow the decision.01 / 04
01 / THE PROBLEM

Start with the friction.

Coding agents can change dependencies, application code, and infrastructure faster than a reviewer can inspect every action. The missing piece is a reviewable boundary between a change and its consequences.

Why it matters

A finding needs context: what was detected, which rule applied, and whether a person should review the change. A risk score alone cannot explain a decision.

02 / WORK & CONTRIBUTION

What the work contains.

My résumé records work on the React dashboard, Express API, deterministic scanners, and policy-decision workflows. Those parts are present in the public repository.

03 / ARCHITECTURE

Give each part a job.

  1. 01

    Inspect the repository

    A repository walker feeds deterministic secret, Dockerfile, Kubernetes, workflow-log, and dependency scanners. Findings retain file-level evidence.

  2. 02

    Evaluate explicit policy

    Ordered, versioned rules match structured findings. Decisions are ALLOW, WARN, REQUIRE_APPROVAL, or BLOCK; fallback rules preserve a defined outcome.

  3. 03

    Keep the explanation

    Prisma models store findings, rule snapshots, decisions, remediation, approvals, and audit events in PostgreSQL. The React dashboard presents that history.

04 / CONSTRAINTS & TRADE-OFFS

Boundaries are part of the design.

Deterministic signatures make decisions explainable, but they do not detect every secret or vulnerability. Dependency inventory is not a full CVE scanner. Authentication and multi-tenant governance remain future work in the README.

05 / VERIFICATION

Evidence before confidence.

Reviewed scanner orchestration, rule matching, decision types, and persistence source at the linked revision. No automated test suite was found in that checkout, and the database-backed system was not run in this portfolio review. The explorer is a labeled explanation of the source, not a live security service.

Evidence reviewed on 18 October 2026 at the source revision linked below. Local checks do not establish production readiness.

06 / PRESENT STATE & NEXT STEPS

There’s a next iteration.

Public control-plane demo. Add scanner regression fixtures, test policy edge cases, and validate the full database-backed workflow before expanding CI integrations.

07 / FOLLOW THE SOURCE

Read past the story.

The source links point to the exact revision reviewed, rather than a moving branch.

Revision c761df8e74fc
ANOTHER QUESTION WORTH EXPLORING

SecondYouth