Lumiq – Multi‑Agent AI Investment Analyst

Abstract

Lumiq is an AI‑powered investment analysis platform that replicates a lean equity research desk using six specialized agents for fundamentals, market data, risk, technicals, financial statements, and final recommendations. Each agent is grounded in a curated 128‑document FinanceBench corpus plus live Yahoo Finance data, which reduces hallucinations and keeps outputs tied to real company and market context. The system returns structured BUY/SELL/HOLD calls with conviction scores, price targets, stop‑loss levels, and timelines, making institutional‑style research accessible to students and retail investors. The full experience is delivered through a custom Gradio UI, deployed as a Hugging Face Space so users can run the multi‑agent workflows directly in the browser.

What I did (methods)

• Multi‑agent architecture: Designed a team of Fundamental, Market Data, Risk, Technical, Financial Statement, and Chief Investment Officer agents, each with dedicated prompts, metrics, and sentiment scoring that mirror real investment analyst roles.
• RAG pipeline: Implemented a two‑stage retrieval system with Sentence Transformers embeddings, FAISS vector search, and cross‑encoder reranking to surface the top 3 most relevant contexts per query from a 128‑document financial knowledge base.
• Live market data: Integrated yfinance to fetch price history, valuation ratios, and sector benchmarks for 10 large‑cap tickers (AAPL, MSFT, NVDA, etc.), which agents use when comparing peers and judging valuation.
• Financial statement parser: Built a document parsing pipeline for PDF, Excel, and CSV filings that extracts key metrics (revenue, income, leverage, margins) and feeds them into a Financial Statement Analyst agent that generates a 6‑section report.
• Web app & deployment: Designed a custom Gradio interface with separate tabs for Stock Analysis and Financial Statement Analysis, wired validation and orchestration into the UI, and deployed the application as a Hugging Face Space using Groq LLM endpoints for low‑latency inference.

Key findings

• Balanced recommendations: The agent team produced more nuanced calls, such as a SNAP case study where bearish fundamentals, risk, and technicals outweighed neutral market data to yield a high‑conviction SELL with an explicit execution plan.
• Grounded reasoning: Combining FinanceBench with live Yahoo Finance feeds led to specific, data‑backed commentary (e.g., debt‑to‑equity vs sector norms, margin gaps vs peers) instead of generic LLM investment advice.
• Statement insights and data issues: The financial statement workflow successfully surfaced profitability strengths and leverage levels while also flagging extraction anomalies (like suspicious revenue fields), encouraging users to treat the analysis as a starting point for deeper due diligence.
• Transparent scoring: A simple scheme where each specialist outputs bullish (1), neutral (0), or bearish (−1) signals, summed into BUY/SELL/HOLD with a 1–10 conviction score, made the system’s reasoning easy to audit and learn from.

What this shows about me

• Comfortable architecting end‑to‑end AI systems that combine RAG, multi‑agent orchestration, and domain‑specific reasoning instead of relying on a single monolithic LLM call.
• Able to work with real‑world financial data—from live market APIs to messy multi‑format SEC‑style filings—and turn them into clean, model‑ready inputs.
• Focused on interpretability and user trust, using explicit scoring rules, grounded retrieval, and explainable reports so non‑experts can understand and critique recommendations.
• Experienced in building, styling, and deploying production‑like applications in Python and Gradio, including validation utilities and orchestration logic for coordinating multiple agents.