TeamAlbertAnna

Project Description (succinct)

RadarAI is a conversational “company intelligence radar” for financial and narrative analysis. Instead of navigating static dashboards, users ask questions in natural language (e.g., “compare Lululemon vs. Inditex” or “why does adidas rank high on narrative momentum?”), and Claude Sonnet 4 generates the appropriate interactive UI in real time using the A2UI (Agent-to-UI) declarative protocol.

There are no pre-built pages or templated views. Each screen is created on demand from a 9-dimension scoring framework (Growth Quality; Revenue Durability; Profitability & Unit Economics; Capital Discipline; Competitive Positioning; Narrative & Tone Momentum; Governance & Alignment; Expectations vs. Reality; Structural Risk Exposure). The client renders Claude’s output as a structured JSON component tree (radar charts, comparison tables, drill-down cards), and the layout adapts to the question.

What’s different about the interface

• Chart-as-navigation: Every data point is clickable; selecting a dimension triggers a new, targeted deep-dive view.
• Annotation mode: Users can mark up a visualization (circle, underline, write notes). The annotated image is sent back as context, and the system responds to what the user highlighted.
• Progressive rendering: UI components stream in as they’re generated, so the analysis appears incrementally rather than behind a single loading state.
• UI validation + repair: A validator catches structural issues (missing references, circular dependencies, schema errors) and auto-fixes them before render.

Data pipeline (high level)

The backend ingests SEC EDGAR 10-K filings and company IR materials, extracts structured financial metrics and text-based signals (e.g., risk-factor expansion, pricing-power language, tone shifts), stores both features and document chunks in RedisVL with Voyage AI embeddings, then normalizes signals into 0–100 peer-group percentiles using robust z-score methods.

---

Technologies, Frameworks, and Libraries

Frontend / Generative UI

• A2UI (v0.8)
• Lit 3.3.1 + @lit-labs/signals
• Chart.js 4.5.1
• Vite 6.0
• TypeScript 5.8.3

AI / Agent Layer

• Claude Sonnet 4 (Anthropic SDK)
• A2UI Agent Framework (Python): schema validation, prompt generation, A2A integration
• Voyage AI embeddings

Data Pipeline & Backend (Python)

• PyMuPDF / pdfplumber / camelot-py (PDF + table extraction)
• BeautifulSoup4 / trafilatura (IR site extraction)
• RedisVL (vector store)
• Pandas / NumPy (processing + normalization)
• Pydantic (validation)
• httpx / tenacity (resilient EDGAR fetching)

Protocol & Infrastructure

• A2A SDK
• Google ADK
• MCP server for exposing A2UI over standardized tool interfaces