Finlet AI: An Evidence-Driven Multi-Agent Financial Intelligence Platform Using Model Context Protocol and Deterministic Risk Governance

Standalone conversational language models struggle to satisfy audit and compliance standards in institutional equity research. When an equity analyst prompts an ungrounded model to evaluate a listed company, three operational defects appear almost immediately: the model quotes outdated share prices from its pre-training cutoff, invents page or footnote numbers when summarizing regulatory filings, and produces inconsistent numerical risk ratings when executed repeatedly on identical inputs. This paper presents Finlet AI, a twelve-layer multi-agent financial intelligence platform that decouples quantitative market arithmetic and primary-source retrieval from neural text synthesis. Once an analytical query is submitted through our asynchronous FastAPI gateway, a stateful LangGraph supervisor triggers concurrent parallel fan-out across three specialized domain workers covering historical price series, live financial news feeds, and statutory annual reports isolated behind standardized Model Context Protocol (MCP) tool interfaces. Technical price action, 20-day volume momentum ratios, and 30-day annualized volatility (σ ) are computed directly from three-month 30d Open-High-Low-Close-Volume (OHLCV) series fetched via Yahoo Finance. Real-time news catalysts are gathered through Google News RSS syndication and sanitized of HTML markup, while balance-sheet debt and liquidity disclosures are extracted page by page from SEC Form 10-K and Indian annual reports indexed locally in ChromaDB using 384-dimensional all-MiniLM-L6-v2 sentence embeddings. Rather than delegating risk scoring to subjective prompt completion, our Risk & Governance module evaluates a bounded mathematical index weighting volatility (45% weight), news sentiment risk (35% weight), and filing flags (20% weight). Live validation on Reliance Industries (RELIANCE.NS) demonstrates real-time tracking at ₹1,184.50 (−2.03%), 23.1% annualized volatility, 1.0x volume momentum, and a deterministic risk score of 0.365 (LOW), while high-risk runs (≥ 0.65) pause at a persistent SQLite Human-in-the-Loop (HITL) checkpoint before compiling the final ReportLab PDF dossier..