DossierMind — Evidence before advice

DossierMind helps a client and an accountant understand a shared document archive without losing the distinction between original evidence and AI interpretation. Its competition demo uses two entirely fabricated client archives and three protected demonstration profiles. Original documents remain encrypted and separately authorized by client. PDF, spreadsheet, CSV, XML and email extraction creates source-linked passages. A clickable brain-shaped map connects documents, wiki proposals and attributed contributions, letting a reviewer return to the underlying source. Server-side AMD inference powers document wiki proposals, scanned-page transcription and an Advisor that separates quotations from interpretations and proposed actions. Quotations are checked against supplied passages. Visual output can be compared with its original page. When evidence is insufficient, the system can abstain rather than fabricate an answer. Draft PDF reports preserve provenance and review limitations. This is a working competition prototype, not certified accounting advice or a commercially finished service. Retrieval is bounded and lexical; exact quotation checks do not establish semantic correctness, payment, tax deductibility or complete coverage. AI results require professional review. The public release contains only demo code and fabricated examples; private business materials, customer archives, credentials and runtime databases are excluded. Preexisting components are disclosed in PROVENANCE.json. AMD integration and subsequent prototype increments are documented through actual repository commits. Track 3 – Reinvent Commerce: DossierMind turns fragmented business records into source-linked insights that help accountants deliver faster, better-informed services to business clients. Its commercial value lies in reducing document-search effort, improving service quality and supporting actionable professional recommendations.