MANAN AGRAWAL. An AI Powered Investment Analytics Platform for Retail Investors: Integrating Quantitative Finance, Real Time Data Processing, and Generative Artificial Intelligence. (Underthe direction of DR. ARINDAM MUKHERJEE.)The democratization of institutional grade financial analytics represents one of the most consequential challenges in modern financial technology. Retail investors historically have operatedunder severe informational disadvantages relative to their institutional counterparts, lacking accessto the sophisticated quantitative tools, real time data pipelines, and expert advisory systems thatgovern professional portfolio management. This thesis presents the design, implementation, andevaluation of an integrated investment analytics platform that bridges this capability gap through theconvergence of quantitative finance, real time data processing, and generative artificial intelligence.The proposed system integrates five tightly coupled subsystems: (1) a machine learning predictionengine comprising a hybrid LSTM XGBoost architecture trained on approximately 50 years ofhistorical equity market data, which generates directional price forecasts and sector level tradingsignals from technical, fundamental, and sentiment features; (2) a multi agent AI decision architecture consisting of a Supervisor Node that orchestrates three specialized agents a Technical AnalysisAgent, a Fundamental Analysis Agent, and a News/Sentiment Agent each producing buy/sell/holdsignals with confidence scores that the Supervisor fuses via weighted voting; (3) a quantitativeanalysis engine that computes risk adjusted performance metrics including the Sharpe ratio, Sortinoratio, Jensen’s alpha, market beta, Value at Risk (VaR), maximum drawdown, and portfolio leveldiversification scores; (4) a multi factor valuation engine that estimates intrinsic asset value andclassifies securities into undervalued, fairly valued, and overvalued categories; and (5) a generativeAI advisory module that deploys a domain-adapted large language model (LLM) served locallyiiiivvia the Ollama inference runtime, fine-tuned on financial reasoning corpora to produce grounded,human-readable buy, sell, and hold recommendations alongside quantitative risk explanations,without reliance on external cloud API services.The system architecture follows a full stack paradigm with a React/Next.js frontend, a Node.jsRESTful backend, an ML inference pipeline, and an asynchronous multi agent reasoning layer.Historical backtesting over a 50-year evaluation horizon demonstrates that the multi agent systemachieves a Sharpe ratio of 1.84 and a cumulative return of 312% versus 187% for a buy and holdbaseline, with a maximum drawdown 14.3 percentage points lower than a single model logisticregression baseline. The ML prediction engine attains 68.4% directional accuracy on out of sampletest data, with a weighted F1-score of 0.71 across Buy/Hold/Sell classes. The integration ofmulti agent signal fusion with quantitative rigor and natural language explainability addresses adocumented gap in the fintech literature: the absence of systems that are simultaneously analyticallysophisticated, explainable, and cognitively accessible to non expert users.This work makes the following principal contributions: (1) a novel hybrid ML–multi agent architecture for AI-augmented retail finance platforms; (2) a formalized quantitative pipeline adapted forreal-time web deployment; (3) an Ollama-based locally-deployed, fine-tuned LLM advisory systemthat operates without external API dependency; (4) a backtesting-validated decision frameworkspanning five decades of market data; and (5) empirical evidence that supervisor-based multi-agentfusion demonstrates consistent improvement over single-model and rule-based financial decisionsystems across all evaluated performance dimensions under the studied market conditions.
