Vault Form Agent

Secure document locker and browser agent for redacted, human-approved form filling across sensitive workflows.

Comment

Inspiration

Filling government, healthcare, insurance, and enterprise forms is still a painful copy-paste workflow. The harder version is not just typing into fields: users need to find the right official site, understand what the form asks for, extract details from scattered documents, protect PII/PHI, avoid accidental submission, and keep an audit trail.

Vault Form Agent explores what a safer form-filling agent should look like when documents are sensitive and the user still needs control.

What it does

Vault Form Agent is an enterprise-grade web app with a secure document locker, redaction policy engine, chat-style form agent, visible browser runner, audit trail, CLI, and Codex/Claude-style skill scaffold.

Users can upload any kind of document with a name, description, tags, metadata, sensitivity label, consent flag, and retention context. The backend extracts text from PDFs, scanned images, text files, JSON, and document-like inputs through a Docling-ready extraction layer with OCR fallbacks. Before data reaches the agent, built-in and custom redaction rules can mask PII/PHI such as emails, phone numbers, SSNs, dates of birth, member IDs, policy numbers, patient IDs, or any custom user-defined field.

The agent page behaves like a chat workspace. A user selects whether the agent should use no documents, selected documents, all non-sensitive documents, or all documents. The user then describes the task in natural language, optionally enables web research, reviews the plan, and lets the browser runner prepare a real web form. The demo proves the process on a public Google Form: eight fields are matched and prepared while a submit guard prevents accidental final submission.

Visual proof

The GitHub README now includes an embedded visual walkthrough so judges and automated reviewers can see the project without hunting through folders:

How it works

  • FastAPI backend manages the locker, extraction, redaction, planning, audit logs, and browser orchestration.
  • React/Vite frontend provides a multi-page enterprise UI: Dashboard, Document Locker, Agent Chat, Redaction Settings, Approvals, and Audit.
  • Redaction engine supports built-in rules plus custom regex, exact-value, dictionary, and metadata-key rules with multiple masking strategies.
  • Browser runner uses labels, nearby context, DOM roles, JavaScript execution, and page analysis to prepare matching fields.
  • Human-control guard rails keep high-risk actions, especially final submit, reviewable.
  • CLI and skill folders make the capability testable as a developer workflow, not just a web app.

How Codex and GPT-5.6 were used

Codex with GPT-5.6 was used for the majority of the project: planning the architecture, designing the enterprise UI, generating and iterating backend services, building the React screens, writing the redaction and document-scope mechanisms, implementing the visible browser fill runner, creating sample data, producing demo recordings, and writing the README and docs.

The key decisions made with Codex were: make the agent chat-first, require document scope selection, separate extraction from redaction, treat sensitive data as policy-controlled, keep submission human-approved, and include a reusable CLI/skill path for Codex-style workflows.

Built during the hackathon

The final repo includes setup instructions, sample data, demo recordings, security notes, architecture documentation, redaction handbook, browser-agent handbook, demo guide, visual walkthrough screenshots, and a hackathon submission sheet for judges.

Demo proof

The included demo assets show the app workflow and a real Google Form preparation run where eight fields are filled and final submission is blocked for safety.

Built With

Share this project:

Updates