DeepDiligence

Overview

DeepDiligence is a multi-agent AI system for investment due diligence, built as the final project for Cornell's ENMGT5400. A user enters a company name; five specialist agents run in parallel; a synthesis agent reconciles their findings; and the result is a structured investment memo with source citations, confidence scores, and conflict detection.

On top of the single-company analysis, it also produces weekly stock rankings across a 95-stock universe and a portfolio builder that allocates capital based on AI ratings, with live prices and S&P 500 benchmarking. This is v1.0 — complete and deployed.

Problem

Investment due diligence means pulling together financials, competitive positioning, team signals, risk factors, and market data from dozens of scattered sources — then reconciling them into a defensible recommendation. It's slow, inconsistent, and hard to audit: when a conclusion is wrong, it's rarely clear which underlying claim drove it.

Solution

A router/aggregator pipeline that fans a request out to five isolated specialist agents and reconciles their structured outputs into an auditable memo:

Parallel specialists — financial, team/culture, market/competitive, risk/sentiment, and a quantitative ML agent run concurrently
Isolation by design — agents cannot talk to each other; each returns a typed sub-report
Reconciliation — a synthesis agent sees only the structured sub-reports (never raw documents), detects contradictions, and scores confidence
Traceability — every claim in the final memo carries at least one source id
Beyond one company — weekly BUY/HOLD/SELL rankings across 95 stocks, plus an AI-rated portfolio builder benchmarked to the S&P 500

Architecture: Router / Aggregator

An orchestrator dispatches five agents concurrently (ThreadPoolExecutor, 5 workers), then gates a synthesis step:

Financial Analyst → SEC EDGAR filings, investor-relations pages
Team & Culture → job postings, leadership signals
Market & Competitive → competitors, TAM, positioning
Risk & Sentiment → news sentiment, litigation, macro flags
Quantitative Momentum → Random Forest ML on 3yr OHLCV + 14 technical indicators

Each agent returns a typed AgentSubReport. The Synthesis Agent merges those into a Pydantic InvestmentMemo — six sections, per-claim source ids, and an overall confidence score. Because agents are fully isolated and synthesis only ever sees structured reports, bias can't propagate and every output stays auditable.

The Quantitative Agent

Unlike the four LLM tool-calling agents, the quant agent is pure ML:

Downloads 3 years of OHLCV via yfinance, then engineers 14 features — RSI, MACD, Bollinger position, MA20/50/200 crossovers, multi-window momentum, ATR, and volume z-score
Trains a RandomForestClassifier (300 trees, depth 6, balanced classes) on an 80/20 time-ordered split — no shuffling, to prevent look-ahead bias
Calibrates confidence conservatively (caps reported confidence at 0.75) and returns a zero-confidence stub for private companies with no ticker

Portfolio Builder

Allocates capital from the latest weekly report:

raw_weight = suggested_weight_pct × rating_boost × confidence
# STRONG BUY = 1.5×, BUY = 1.0×

Weights are then capped and iteratively re-normalized (excess redistributed from capped positions), live prices are fetched per request via yfinance, and results are benchmarked against the S&P 500 across 1W/1M/3M/YTD.

Tech Stack

Backend: Python 3.11, FastAPI + uvicorn, Server-Sent Events for real-time agent progress
LLM: Anthropic Claude API (claude-sonnet-4-20250514)
ML: scikit-learn RandomForestClassifier, numpy
Data: yfinance, SEC EDGAR, NewsAPI, BeautifulSoup scraping
Frontend: React + TypeScript, TanStack Router, Tailwind CSS
Contracts: Pydantic models for every inter-agent boundary
Caching: disk-based, SHA256-keyed JSON per agent per company