Ten Institutional
Minds. One
Portfolio.
Simulates Goldman Sachs, Bridgewater, JPMorgan, and 7 more analyst personas debating your holdings across three rounds of Delphi reasoning, then synthesizes a CIO-level recommendation.
The Problem
One model, one perspective. That's not how investment decisions actually get made.
When you ask a single LLM to analyse your portfolio, you get one best guess. Even with live data, the model has no mechanism to challenge its own reasoning. The result is confident mediocrity: authoritative-sounding output that anchors on its first inference and never revises.
The Fix
A better architecture, not a better prompt.
Real investment decisions come from debate: analysts challenging each other's assumptions, risk managers pushing back on growth theses, quants flagging reversed momentum signals. The final recommendation is the synthesis of disagreement, not the output of a single mind.
That's what Project Delphy builds.
Architecture
The Delphi Method, applied to AI.
Each round, analysts see each other's reasoning and revise. Independent experts converging through structured debate: the same method used to aggregate expert judgment since the 1950s.
Round One
Independent Analysis
All 10 analyst agents run in parallel. Each receives live market data: prices, yields, earnings, analyst ratings via the real-time RAG pipeline and issues an independent assessment of your portfolio.
Round Two
Cross-Validation
Each analyst reads the other nine analyses and updates their view. Overconfident positions get challenged. Overlooked risks surface. This mirrors how a real investment committee works before a vote.
Round Three
Convergence
Remaining disagreements from Round 2 are addressed. Analysts settle on their final, post-debate positions, having integrated macro, fundamental, quant, and strategic perspectives.
Synthesis
Investment Committee Report
A meta-agent acting as Chief Investment Officer reads the full analysis output and synthesises a final recommendation: consensus picks, dissent flags, rebalancing action items, and macro context.
Live Market Data, Before Every Run
Before Round 1 begins, the system fetches current prices, S&P 500 / Nasdaq / VIX, treasury yields, top analyst ratings, earnings calendar, and market-moving headlines, giving every analyst the same live context.
The Committee
Ten genuinely different investment philosophies.
Not ten slightly-different clones. Bridgewater and Citadel will frequently disagree. That's the point. The synthesis of disagreement is more valuable than artificial consensus.
Goldman Sachs
Macro Strategist
Macro-driven, risk-adjusted returns. Focuses on top-down sector rotation and rate-sensitive positioning.
Morgan Stanley
Equity Analyst
Fundamental equity analysis through DCF valuation. Emphasis on earnings quality and competitive moats.
Bridgewater
Risk Parity Analyst
All-weather, macro risk parity. Balances across economic environments: growth, recession, inflation, deflation.
JPMorgan
Portfolio Strategist
Diversification and downside protection. Stress-tests portfolios against tail scenarios and liquidity events.
BlackRock
Factor Analyst
Factor-based, systematic investing. Analyses exposure to value, quality, momentum, low-volatility, and size.
Citadel
Quant Analyst
Quantitative, alpha-seeking. Identifies statistical mispricings and evaluates signal persistence.
Harvard Endowment
Long-Horizon Allocator
Long-horizon, alternatives-heavy. Evaluates portfolios for real assets, private equity, and illiquidity premium.
Bain & Co.
Strategic Analyst
Strategic business and value lens. Assesses competitive position, management quality, and value creation.
Renaissance Technologies
Statistical Analyst
Statistical pattern recognition in price and volume data. Evaluates momentum, mean-reversion, and anomaly signals.
McKinsey
Operational Analyst
Operational value and portfolio strategy. Examines industry dynamics, operational efficiency, and strategic fit.
33 API calls per Delphy run: 3 rounds across all analysts, synthesised by a CIO meta-agent
Capabilities
Built for real portfolio workflows.
Multi-Format File Ingestion
Upload Vanguard QFX/OFX brokerage statements, Bank of America TXT exports, PDF documents, Excel spreadsheets, or images. The system parses and extracts portfolio context automatically.
Real-Time RAG Pipeline
Live prices, S&P 500 / Nasdaq / VIX, treasury yields, sector ETF movers, earnings calendar, and analyst ratings fetched fresh before every Delphy run via Claude's web search tool.
Three-Round Delphi Reasoning
Structured on the Delphi Method: Round 1 collects independent expert opinions in parallel, Rounds 2 and 3 expose each analyst to the others and allow revision, the CIO meta-agent synthesizes final consensus.
CIO Synthesis Report
The meta-agent Investment Committee Report delivers consensus picks, explicit dissent flags where analysts disagreed, actionable rebalancing suggestions, and macro risk context.
Session Persistence
Portfolio context, analysis results, and conversation history auto-save and restore between sessions. Pick up your analysis exactly where you left off.
Bilingual Interface
Full English / Korean bilingual UI: toggle between languages for the full portfolio analysis, chat, and report output. Designed for international users.
Daily Market Report
A standalone script generates a comprehensive daily briefing: 8 market sections including sector performance, top movers, macro indicators, economic calendar, and analyst consensus digest.
Contextual Chat
After the analysis runs, continue the conversation with full portfolio context. Ask follow-up questions, drill into analyst disagreements, or request scenario analysis.
Daily Report
A complete market briefing, every morning.
The standalone daily_report.py script uses Claude and live web search to generate an eight-section briefing covering the full market picture (macro, sectors, movers, calendar, and analyst consensus), saved as a dated Markdown file.
Schedule it with cron or run it manually. The full report takes roughly 60–90 seconds and costs approximately $0.05–0.10 per run.
See the script on GitHubReport Sections
- 01 Market Overview: S&P 500, Nasdaq, Dow, Russell 2000, VIX, treasuries, DXY
- 02 Sector Performance: All 11 S&P 500 sector ETFs with daily change
- 03 Top Movers & Headlines: 5 gainers, 5 losers, 5 market-moving stories
- 04 Portfolio Recommendations: 3 actionable picks with thesis, target, risks
- 05 Watchlist: 5 stocks to monitor with entry criteria
- 06 Macro & Economic Indicators: Fed rate, yield spread, CPI, PCE, unemployment
- 07 Economic Calendar: Upcoming Fed speeches, earnings, data releases
- 08 Analyst Consensus Digest: Upgrades, downgrades, and price target changes
$ python daily_report.py
Fetching live market data...
Generating Section 1: Market Overview
Generating Section 2: Sector Performance
...
Report saved → reports/2026-04-02.md
✓ Complete · 8 sections · ~$0.08
Stack
Intentionally simple infrastructure.
The backend is a straightforward Express server with 6 endpoints. All the intelligence lives in the prompt engineering and the orchestration logic that sequences the 33 API calls.
~$0.10–0.30 per Delphy run (33 API calls) · ~$0.05–0.10 per daily report
Project in Two Parts
Built with two lenses: technical and strategic.
How Project Delphy Was Built
A technical account of the multi-agent architecture: how the Delphi Method was applied to 10 analyst personas, why 3 rounds of cross-validation produce better recommendations than any single agent, and what the implementation looked like end-to-end.
"Real investment committees argue. I wanted to see what happens when AI agents do the same."
What If This Were a Product?
A first-person case study on trying to turn a technically impressive build into a viable product. I explored every path seriously: consumer SaaS, B2B, platform licensing. Then the numbers did what numbers do.
Good business judgment includes knowing when not to build, not just when to start.
Open Source
Try it with your own portfolio. All you need is an Anthropic API key.
A full Delphy run costs about $0.15 on average. The full source code, deployment guide, and documentation are on GitHub.