Can I vibecode Consensus?

Build a sophisticated personal academic research assistant inspired by Consensus.
The stack MUST be Next.js 15 (App Router), TypeScript, Tailwind CSS, SQLite, and an OpenAI-compatible API client for LLM features (designed to work with local models via Ollama or remote APIs).

Core Objective: Build a private AI-powered search and synthesis workspace that processes user-provided academic papers (PDFs or text), extracts metadata, and answers complex research questions with evidence-based citations.

Key Features to Implement:
1. Document Ingestion & Metadata Extraction:
   - Users can upload academic papers.
   - Automatically extract study metadata (study design, sample size, population, duration) using an LLM.
   - Store documents and metadata in SQLite.

2. AI-Powered Search & Synthesis:
   - Implement a search bar that takes natural language queries.
   - Retrieve relevant documents from the local database.
   - Synthesize a cohesive summary answering the query, with inline citations linked directly to the uploaded source documents.

3. The "Consensus Meter":
   - Analyze the retrieved search results to determine the level of scientific agreement on a "Yes/No" or "Effect/No Effect" question.
   - Display a visual indicator summarizing whether the local literature agrees, disagrees, or is mixed.

4. Study Snapshots & Advanced Filtering:
   - Display search results as cards containing the extracted metadata (methodology, sample size, etc.).
   - Allow filtering results by study design, year, or custom tags.

5. "Ask Paper" Interactive Chat:
   - Provide a chat interface for individual papers, allowing users to ask deep questions about a specific document.
   - The AI must highlight or reference exactly where the information is located in the text.

6. Citation Management:
   - Generate standard citations (APA, MLA) for any document in the workspace.
   - Allow exporting citations and saved lists.

Technical Constraints & Polish:
- Start from an empty folder.
- Data must be stored locally in SQLite.
- Do NOT include analytics, tracking, or telemetry.
- Provide a clear .env.example for API keys.
- Build a polished, accessible UI with clear loading states for AI operations, empty states for a fresh workspace, and error handling for failed extractions.
- Ensure the interface feels premium and responsive.
- Include unit tests for the metadata extraction logic and an end-to-end smoke test for the core ingestion and search loop.
- Provide a README explaining the architecture and setup.