Prophet

Prophet is a personal investment intelligence system that autonomously researches companies, markets, events, sources, and portfolio exposures; maintains a persistent, time-aware model of evidence and views; continuously identifies what changed and what matters; challenges and verifies its own conclusions; tests ideas through simulation without real trading; and learns from later outcomes.

It’s designed to give one investor the research depth, institutional memory, portfolio context, skepticism, and continuous monitoring of a small investment firm—without placing trades or moving money.

Unlike a generic chat assistant, Prophet maintains research state across sessions. Unlike a broker, it never places real orders. The LLM organizes and challenges evidence; deterministic services own portfolio truth, calculations, timestamps, and guarded state transitions. Private model chain-of-thought is not exposed.

What Prophet Does

• Reconstructs holdings from deterministic transaction records and optional broker-confirmation ingestion.
• Preserves research evidence with source lineage and event, publication, and ingestion times.
• Maintains one current accepted view per subject, separate from the evidence, analyses, and unresolved questions that support or challenge it.
• Connects holdings to fundamentals, investor expectations, peers, themes, historical analogies, contradictions, and portfolio-level transmission paths.
• Runs bounded research, verification, monitoring, reminders, integrity review, and source-learning workflows in the background.
• Can deliver opt-in, generic owner notifications when a durable watch or reminder changes state; the full evidence and action plan remain inside Prophet.
• Surveys an operator-defined investable universe assembled manually or from tracked positions, researched entities, and latest stored benchmark snapshots, for provisional opportunities, then evaluates immutable discovery-time hypotheses against dated market and cash controls, exposing coverage, assumptions, evidence, skips, failures, and provider cost.
• Tests ideas through shadow experiments and paper accounts, then records outcomes and lessons without sending real orders.

How It Works

flowchart LR
    B["Broker activity<br/>Gmail / CSV / manual / optional Plaid"] --> P["Portfolio ledger<br/>holdings / cash / transactions"]
    R["Research<br/>web / filings / YouTube / market data"] --> E["Evidence store<br/>source / date / provenance"]
    P --> X["Prophet research context"]
    E --> X
    X --> L["LLM research and challenge"]
    L --> D["Guarded state updates<br/>accepted views / questions / risks / catalysts"]
    D --> U["Portfolio / Research / Knowledge / Review"]
    D --> M["Watches / reminders / shadow experiments"]
    M --> E

Loading

The model can organize evidence, propose hypotheses, and identify weak points. It cannot directly rewrite portfolio truth or promote unsupported claims into accepted state.

Experimental software: Prophet is not investment advice. Its data and analysis may be incomplete or incorrect; verify important evidence and decisions independently.

Run Locally

Requirements:

• Python 3.11-3.14
• Poetry 2.3.2, installed in its own environment (prefer pipx install poetry==2.3.2)
• Node.js 24 LTS and npm
• Docker with Compose, or an existing reachable PostgreSQL 16 server

Start Prophet from the repository root:

python scripts/prophet.py

On Windows, use py -3.11 scripts\prophet.py. The same command handles first run and ordinary startup: it checks prerequisites, creates a private .env with a generated local database password when needed, reuses the configured PostgreSQL server or starts the local Compose service when none is listening, installs lockfile-pinned dependencies, applies migrations, rebuilds changed frontend code, waits for readiness, and opens Prophet. It never installs or starts Ollama or another model provider.

Keep the launcher window open while using Prophet. Press Ctrl+C there to stop the backend and frontend processes it started; PostgreSQL remains available for the next run. Diagnose setup or inspect a running stack with:

python scripts/prophet.py doctor
python scripts/prophet.py status

Use python scripts/prophet.py --dev for frontend hot reload. The launcher will not kill an unknown process when ports 3000 or 8000 are occupied; it reports the conflict so the owning process can be handled deliberately. Logs live in the ignored .prophet-local/runtime/ directory.

The production frontend includes an installable web-app manifest. On a secure origin, use the browser's Add to Home Screen or Install app action to run Prophet in a standalone window. Offline navigation shows an explicit unavailable state; Prophet does not cache portfolio, research, chat, or API data for offline use. The default launcher remains loopback-only, so installation on another device requires a separately secured private HTTPS access path rather than exposing ports 3000 or 8000 to the LAN or public internet. See Private access for the supported Tailscale Serve boundary, exact operator-identity gate, mobile installation, notifications, configuration, and threat model.

Do not install Poetry into backend/.venv; Poetry must remain isolated from the environment it manages. The older dev_up.sh, stable_up.sh, and status commands remain as macOS/Linux compatibility aliases, but delegate to the same launcher rather than maintaining a second service-orchestration path.

The API health endpoint is http://127.0.0.1:8000/health. If the database port is already in use, set another loopback POSTGRES_PORT in the private .env before retrying.

Configure model and research integrations from Settings > Research. Prophet's implemented LLM provider registry currently contains:

• NVIDIA NIM (Cloud): uses an API key and supports a configurable hosted model and base URL. Streaming is supported.
• Codex CLI: uses an installed and authenticated Codex CLI. Prophet does not manage an API key, model, or base URL for this adapter. Streaming is not supported.
• Ollama (Local): supports a configurable local model and base URL, plus streaming. It remains unavailable unless the operator explicitly enables local-provider use with LLM_ALLOW_LOCAL_PROVIDER=true; Prophet never starts Ollama.

External source discovery is separate from the LLM and from Prophet's stored retrieval system. Prophet can query an operator-supplied SearXNG endpoint first and use Tavily as a metered fallback. Search results identify candidate pages; Prophet attempts to fetch the underlying page through its outbound URL safety policy before creating evidence. Prophet neither installs nor starts SearXNG. Tavily is a search provider, not an LLM provider or RAG system.

The settings UI stores configured secrets in an ignored mode-0600 sidecar and only returns key-presence flags. Environment-based deployments may instead set NVIDIA_API_KEY and TAVILY_API_KEY. SearXNG endpoint and provider order are runtime settings; an optional local Tavily credit budget can bound fallback use.

YouTube ingestion reads an individual video's existing captions first, then assesses what material questions remain unresolved. A bounded deeper pass can verify a specific claim through the normal research pipeline or, when explicitly enabled, use separately installed yt-dlp, ffmpeg, and the OpenAI Whisper CLI for missing or materially incomplete speech. A separately enabled Tesseract adapter can sample bounded frames when on-screen text is material. Raw media stays in disposable temporary workspaces and is removed after extraction. Neither OCR nor speech transcription implies complete visual understanding.

With yt-dlp installed, automation periodically reviews operator-trusted YouTube channel sources. New-upload metadata is first stored as a provisional discovery observation, not evidence. A bounded number of unseen uploads then enter the same caption-first pipeline and remain attributed to the tracked channel. Channel review and automatic ingestion are independently bounded and configurable in .env.example.

Verification

Run the complete local check suite

cd backend
poetry run coverage run --branch --source=investos -m pytest -q
poetry run coverage report --precision=1 --fail-under=40
poetry run alembic check
poetry run black --check investos tests ../scripts
poetry run isort --check-only investos tests ../scripts
poetry run flake8 investos tests ../scripts --select=E9,F63,F7,F82
poetry run bandit -q -r investos -ll
poetry run python ../scripts/repository_policy_check.py
poetry run python ../scripts/audit_python_lock.py
poetry run python ../scripts/audit_state.py

cd ../frontend
npm run lint
npm run build
npm audit --audit-level=low

The backend suite includes synthetic characterization tests for high-risk user-visible behavior. Credentialed Gmail/Plaid flows and browser visual regressions still require live external verification.

Data and Secrets

Do not commit .env, data/, backups, runtime settings, secret sidecars, mailbox data, broker statements, media workspaces, local agent or tool state, or logs. The root .gitignore and .dockerignore exclude these categories. The ongoing repository-policy check rejects private runtime or machine-local artifacts that are eligible for commit. Gitleaks performs credential scanning in CI across the repository history. Rotate any credential that was exposed outside its intended secret store.

The application is local-first and has no multi-user authentication boundary. Do not bind it to a public interface or deploy it as a shared service without adding authentication, authorization, CSRF protection, and a deployment threat model.

Documentation

• Architecture: current components, data flow, and invariants
• Limitations: important operational and quality boundaries
• GitHub Issues: reproducible bugs and proposed improvements

The Python package retains the historical name investos; the product name is Prophet.

Source Availability

Copyright (c) 2026 Eric Wang. All rights reserved. Prophet is a personal, source-available project, not an open-source product or commercial service. You may inspect, clone, run, and modify it locally for personal, non-commercial evaluation, study, and experimentation.

Commercial use, redistribution, resale, sublicensing, hosted services, and incorporation into other products require separate written permission. See LICENSE.md for the complete terms, CONTRIBUTING.md before proposing changes, and SECURITY.md for security reporting and runtime boundaries.