The agent behind VYNN AI

A personal, trustworthy AI financial analyst. Ask about a company in plain words, in any language. The agent decides what the question needs, and code computes every number it gives you. By Zanwen Fu, founder of VYNN AI.

     

Overview

• One loop, any question. A reasoning agent reads the request and picks the tools it needs, or none. There is no fixed pipeline and no intent menu.
• A full analysis in about two minutes. Financials, a ten-tab Excel model with live formulas, screened news and a cited report, usually twenty pages or more.
• Code owns every number. A deterministic calculator sets the rating and the target. A validator checks every figure against it, and requires at least 95% of material sentences to cite the run's evidence.
• Flagged, not hidden. When well-covered analysts disagree with the model, VYNN still answers, with both positions side by side. The call is yours.

What you can ask

You ask	The agent
"Analyze NVDA. Should I buy?"	Pulls the financials, builds a 10-tab valuation model, screens the news and writes a cited report, then answers
"分析英伟达，用中文写报告"	Runs the same analysis and writes the report in Chinese
"How would falling rates hit US banks?"	Reasons from live macro data. No full analysis, answer in seconds
"Flag breakdowns on NVDA and AAPL"	Computes technicals for both and gives the actual levels
"Price a 30-day NVDA 150 call"	Black-Scholes value with delta, gamma, theta and vega
"Best Sharpe weighting for AAPL, MSFT, NVDA?"	Solves the max-Sharpe portfolio and explains the trade-offs
"Odds of a Fed rate cut?"	Reads the market-implied probability from prediction markets
"What's the outlook for Bitcoin?"	A live crypto snapshot, and no DCF: a coin has no cash flows

The heavy path, financials to model to news to report, runs only when a question needs it, and takes about two minutes. Everything else answers in seconds.

How it works

The agent runs one reasoning loop over its tools. A quick question stops there. A question that needs a full analysis goes through the pipeline, and everything it publishes passes the guardrails first. The numbers on the figure are the steps below.

1. Ask. Your question, in any language, reaches the reasoning agent with the conversation so far. History and every tool result enter as data, never as instructions, so a headline saying "ignore your rules" is analyzed, not obeyed.
2. Quick answer. Prices, technicals, options, portfolio risk, prediction-market odds, macro data, funds and crypto come back in seconds, without the full analysis.
3. Full analysis. When the question needs it, the agent gathers the financials, builds the ten-tab model and screens the news in parallel, then writes the report: about two minutes in all.
4. Check. Every figure and claim passes the guardrails. The calculator sets the rating from the gap to price and makes the fair value the 12-month target. The validator corrects any figure that differs and rewrites text that does not cite the run's evidence. The confidence check compares the value with analysts' targets, which never enter the valuation itself.
5. Answer. The answer arrives in the chat, on the report's first page and in the workbook, held to one set of numbers. When analysts disagree, it says so and shows both positions.

The tools

The four analysis stages are tools over one shared state, so a full run keeps its order while independent work runs in parallel. Every tool registers itself with schemas for both OpenAI and Anthropic, so the same loop runs on either provider.

Group	Tools
Data	resolve_symbol (any-language name to ticker), get_prices, get_technicals, get_global_news, get_macro (FRED)
Analysis	get_financials, build_model, analyze_news, write_report, read_report, compare_tickers
Funds and crypto	get_fund (fees, holdings, allocation; never a DCF), get_crypto
Markets	price_option, compute_risk_metrics, optimize_portfolio, get_prediction_markets (Polymarket)
Interface	show_chart, an interactive price chart drawn in the chat

get_macro needs a free FRED key; without one, the agent is simply not offered that tool.

Guardrails

The calculator, the validator and the confidence check are on the figure. Beneath them:

• The workbook is the source of truth. All formulas are live: change one assumption and the projections, both valuations, the sensitivity tables and the summary recompute.
• Two methods that agree by construction. Perpetual growth and exit multiple are reconciled so they describe the same future. The spread that remains is classified, and a method that fails, such as a negative value for a cash-burning company, is reported as failed, not averaged in.
• Inputs from published data. The discount rate is built from the government yield in the currency of the cash flows, Damodaran's equity risk premium and a regressed beta. Every input is printed with its source.
• Checked every night. A canary rebuilds a fixed basket of companies with the released engine. A release ships only with zero unexplained differences from the expected outcomes.
• Untrusted text stays data. News articles, search results and replayed history enter the context marked as data, never as instructions.

The details, from cost of capital to the calibration benchmark, are in the engineering notes.

Real output

Five research records from the current engine, each with its full report and model:

Company	Run	Fair value	Price	Rating
Microsoft	30 Sep 2026	$595.14	$508.96	Buy
NVIDIA	18 Oct 2026	$364.34	$214.13	Strong Buy
Alphabet	18 Oct 2026	$433.19	$340.59	Buy
Visa	18 Oct 2026	$390.24	$402.61	Hold
Apple	18 Oct 2026	$232.21	$343.46	Strong Sell, flagged

Each record states its confidence and its sources. Microsoft's report and model are also in samples/. Research, not investment advice.

Quickstart

Requires Python 3.11 and an OpenAI API key.

git clone https://github.com/Agentic-Analyst/stock-analyst.git
cd stock-analyst
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.lock
cp .env.example .env          # set OPENAI_API_KEY

python main.py --pipeline chat --email [email protected] --timestamp 20260101_120000 \
  --user-prompt "Analyze NVDA. Should I buy?"

Continue a conversation with --session-id. For scripted runs without the agent, pass --ticker and one of the pipelines: comprehensive, financial-statements, financial-model, search-news, screen-news, company-daily-report or sector-daily-report.

Each run writes to data/<email>/<TICKER>/<timestamp>/:

financials/   statements and market data, as collected
models/       the Excel model and its computed values
searched/     news found for the run
screened/     catalysts and risks, with quotes and sources
reports/      the research report
answer.md     the answer
info.log      the full run log

With Docker:

docker build --target production -t stock-analyst .
docker run --rm --env-file .env -v "$(pwd)/data:/data" stock-analyst \
  --pipeline chat --email [email protected] --timestamp 20260101_120000 --user-prompt "Analyze NVDA"

Configuration

Only OPENAI_API_KEY is required. .env.example documents every setting.

Variable	What it enables
OPENAI_API_KEY	The default provider (gpt-6-luna)
ANTHROPIC_API_KEY	Claude models, as an alternative to OpenAI
CHAT_MODEL, --llm	Model choice; python main.py --list-llms lists them
SERPAPI_API_KEY	Richer news discovery
FRED_API_KEY	The get_macro tool (free key)
MONGO_URI, MONGO_DB	Article cache and conversation memory
BENZINGA_API_KEY, FINNHUB_API_KEY	Licensed analyst consensus, used as a benchmark and never as a valuation input

Project layout

main.py                 command line entry point
src/agents/
  generalist_agent.py   the reasoning loop
  tools/                the tools, one module per group
  fm/                   the DCF engine, one builder per workbook tab
  news/                 daily intelligence reports
src/llms/               one interface over OpenAI and Anthropic, with retries and a circuit breaker
src/recommendation_*    the deterministic calculator and the validator
src/confidence_alert.py the flag shown when analysts disagree
prompts/                34 prompt templates, versioned as Markdown
scripts/                the nightly valuation canary and its expectations
tests/                  the test suite CI runs on every change

Tests

pip install -r requirements-test.lock
python -m pytest tests/ -q -p no:cacheprovider

CI runs the same suite on every pull request into main and every push to it.

Limitations

• A DCF does not fit every company. Pre-revenue and deeply cash-negative businesses get a range and a caveat, not a price target.
• Calibration is not yet an accuracy claim. That needs a versioned cohort and a real 12-month outcome backtest, which the benchmark supports but does not yet have.
• News lags. Google News results can trail breaking news by 15 to 30 minutes, so this is not an intraday signal.
• Yahoo Finance throttles bursts. Requests retry with backoff but are not queued across simultaneous analyses.
• Obscure names may need the ticker. Names in other scripts are resolved by transliteration plus search.

Contributing, security, license

Issues and pull requests are welcome. Tools, workbook tabs and prompts extend independently: a new tool is one self-registering module. See SECURITY.md to report a vulnerability.

The source is available to read and evaluate. Use, copying, modification or distribution requires written permission from VYNN AI; see LICENSE. Contributions are assigned to VYNN AI under its terms.