Product
Roadmap
Thesis AI is being built in four phases toward a single destination: an AI-native investment intelligence system that powers both consumer research and institutional capital allocation from the same kernel.
Strategic Frame
The consumer app is distribution and a feedback loop. The kernel is the moat. Every phase adds a layer to the same architecture — nothing gets thrown away, everything compounds.
Four Phases
Phase 1 — Intelligence Engine
Consumer app as research surface. Kernel as foundation.
- iOS TestFlight distribution for early access investors
- Multi-agent research kernel: Macro, Fundamentals, News, Price, Portfolio agents
- InvestmentManager orchestrator with intent-based routing
- Tool interface layer: MarketDataTool, MacroDataTool
- Real-time AI chat with streaming responses via Claude
- Daily market brief generation
- Portfolio and watchlist management
- Thesis cards with confidence scoring
- Internal research console at /console for agent inspection
Phase 2 — Research Platform
Persistent memory. Real-time signals. Power user surface.
- pgvector memory backend — RAG over research history and earnings transcripts
- Live portfolio sync and real-time market data during trading hours
- Push notifications when the thesis on a holding materially shifts
- Persistent conversation memory across sessions
- Saved research, thesis card history, and search
- Confidence calibration tuned on beta feedback
- Thesis Pro desktop client (lightweight, kernel-native)
- Android access
Phase 3 — Fund Infrastructure
Strategy layer. Risk engine. Systematic signal pipeline.
- BaseStrategy layer: regime-conditional signal weighting
- RiskEngine: VaR constraints, Kelly sizing, concentration limits, drawdown guards
- Regime detection: HMM-based macro regime classification
- Factor model integration: momentum, value, quality scoring
- Bayesian signal updating — posterior conviction from agent evidence
- Graph database for cross-asset relationship and sector mapping
- Automated research memos (PDF) with full agent attribution
- Compliance-grade audit log of all agent decisions and data sources
Phase 4 — AI-Native Hedge Fund
Full allocation pipeline. Institutional infrastructure. Developer API.
- Automated signal scoring → risk-adjusted position sizing → execution integration
- Capital allocation dashboard with full pipeline visibility
- Developer API: expose the kernel to third-party integrations and white-label use
- Institutional dashboard for external capital and reporting
- Data partnerships: alternative data ingestion (satellite, web scraping, filings)
- Custom agent configuration and strategy module builder
- Webhooks for real-time thesis event delivery
Why This Architecture Works
The kernel was designed from day one to support all four phases without an architectural rewrite. Key decisions that enable this:
- Transport-agnostic kernel — InvestmentManager takes an AgentContext and returns a dict. It has no knowledge of HTTP, mobile, or UI. Plugging it into a new surface requires zero kernel changes.
- Tool interface layer — agents never call market data services directly. Swapping a data vendor means updating one tool class, not every agent.
- pgvector in the schema from day one — the database already supports semantic vector search. When the memory backend is wired, RAG is additive — not a migration.
- Strategy / Risk as interfaces — the BaseStrategy and RiskEngine abstractions are defined but unimplemented. Any strategy can be added without touching agent logic. The architecture anticipates the fund layer.
- Stateless orchestrator — context is passed explicitly per request. The InvestmentManager is horizontally scalable without shared state.
- LLM abstraction — Claude / Ollama fallback means the LLM backend can be upgraded or replaced without touching agent prompts.
API Access
Coming in Phase 4
The developer API will expose the kernel's agent pipeline to third-party applications, institutional dashboards, and custom integrations. If you are interested in early access, reach out.