ScopeGuard AI

Inspiration

ScopeGuard AI helps agencies, development teams, production teams, and operations managers understand what a client actually requested.

Project requirements often arrive through briefs, emails, meetings, chat messages, and later change requests. These sources can conflict, omit important decisions, or introduce extra work that was never included in the original agreement.

What it does

ScopeGuard AI brings those sources together and analyzes them as one evidence set.

It identifies confirmed requirements, contradictions, missing decisions, assumptions, potential scope creep, risks, dependencies, delivery tasks, and acceptance criteria.

Every important conclusion includes traceable source evidence, such as the original BASELINE or a specific source ID.

The user remains in control throughout the process. AI-generated findings can be approved, rejected, edited, or marked for clarification. ScopeGuard AI does not approve contracts, contact clients, assign work, or finalize scope automatically.

The final approved result can be exported as a project brief, client-facing summary, internal checklist, Markdown, or JSON.

Why AI is needed

This problem requires semantic comparison across multiple sources, contradiction detection, contextual reasoning, uncertainty handling, and transformation of informal instructions into structured project requirements. Keyword rules alone cannot reliably handle those tasks.

How Codex and GPT-5.6 were used

Codex using GPT-5.6 helped build the Next.js architecture, strict TypeScript and Zod schemas, OpenAI integration, evidence validation, review workflow, exports, responsive design, testing, documentation, and deployment preparation.

Human decisions defined the product direction, evidence-first design, human approval requirements, system boundaries, responsible AI safeguards, and final user experience.

Responsible AI

Source traceability
Confidence labels
Human approval before finalization
Prompt-injection resistance
Strict allowed evidence IDs
No automatic client communication
No automatic work assignment
Clear limitations and uncertainty
Demo mode that works without live API access

Testing and reliability

13 automated tests
ESLint passes
TypeScript passes
Production build passes
Real production analysis verified
Secure server-side API key
GitHub Actions quality checks
Responsive desktop and mobile layouts

Challenges

The main technical challenge was maintaining evidence integrity across complex structured model output. A production test exposed a baseline citation issue, which was fixed by introducing the explicit BASELINE evidence ID and restricting model citations to exact allowed source IDs.

Another challenge was balancing reliable structured analysis with production latency. Duplicate retries were removed and the request timeout was adjusted to support a single validated structured response.