Project Overview
PitchProof guides a PR consultant through a structured six-step brief for a product launch or funding announcement, checks every answer for meaning and evidence, exposes the claims that could be repeated publicly, requires human approval, and generates pitches from approved facts only.
Business Problem
Generic AI writing tools turn a thin announcement into fluent but unsupported copy. For PR, a confident sentence that isn't true is worse than no sentence.
Proposed Solution
Deterministic preflight rules check completeness, Gemini reviews every answer and extracts a claim inventory, a human approves, edits or rejects each claim, and three story angles and pitches are generated from approved claim IDs only.
System Architecture
Next-compatible TypeScript app with server-side API routes for review and generation; API keys never reach the browser. Zod validates every model response, and a deterministic demo mode gives reliable presentations without calling the model.
Technical Challenges
Enforcing grounding in code rather than in the prompt: every generated sentence must reference an approved claim ID, and a single unknown or rejected reference rejects the whole response. Provider failures reveal no partial output.
Lessons Learned
Model output is untrusted input. Treating it with the same validation as user input, and making the verification boundary visible to the user, is what makes an AI tool safe to rely on.
Project Gallery
