FinanceAI Agents
Stock Pitch Multi-Agent System
5 orchestrated AI agents produce equity research pitches, gated by up to 12 deterministic checks before anything ships.
- 5
- specialized agents
- ≤12
- validation gates
- 6
- live data sources
- 0
- LLM arithmetic
Architecture
Ticker
Phase 1 · parallel∥ concurrent
Business Intelyfinance · Finnhub
SEC FinancialsEDGAR XBRL
News & SentimentFinnhub
Phase 2
Valuation EngineDCF + comps · Python
Gate
≤12 deterministic checks
Phase 3
Synthesis → DeckBUY / HOLD / SELL · Gamma
InputAgentEngineGateOutput
Overview
A parallel multi-agent system that generates equity research stock pitches. An orchestrator acting as a 'Managing Director1' resolves dependencies and spawns 5 agents in phases: business intelligence, SEC financials, and news/sentiment run in parallel, feeding a valuation agent (deterministic DCF + comparable-company models in Python), which is validated by up to 12 deterministic checks2 before a synthesis agent produces a BUY/HOLD/SELL pitch deck via Gamma. Every data point cites its source; all financial computation is done in Python, never estimated in prose.
- 1Managing Director · The orchestrating agent: it assigns work to the analyst agents and resolves what depends on what, the way an MD runs a deal team.
- 2deterministic checks · Ratios recomputed in Python, sources present, numbers reconciled: every check must pass before a deck is produced.
Highlights
- Dependency-driven orchestration across 5 agents (hard + soft deps, phased execution)
- Deterministic DCF and comps engine with no LLM arithmetic; auto-detects company type
- SEC EDGAR XBRL parsing for audited 10-K / 10-Q / DEF 14A financials
- Cross-agent Redis cache with 3 TTL tiers + Finnhub rate-limit budgeting
- Up to 12 deterministic validation checks gate the deck; run versioning via DVC
Stack & tools
- Claude Code
- Python
- DCF Engine
- SEC EDGAR (XBRL)
- yfinance
- Finnhub
- FRED API
- Upstash Redis
- Gamma
- DVC
- Pytest
- React / Vite