A local, AI-native investment operations workspace for one individual investor. FundOps turns plain-English strategy into a versioned Constitution, runs a durable research pipeline over retained market data, keeps source-backed evidence for every claim, and routes decisions back to you in an Inbox — without ever delegating the investment decision to software.
| Home command surface | Durable workflow runs |
|---|---|
 |
 |
| Markets research hub | Operations settings |
|---|---|
 |
 |
| Screener results | Example investment memo |
|---|---|
 |
 |
| Company page workflow map | Market output dashboard |
|---|---|
 |
 |
FundOps is an investment learning partner, not an autonomous fund manager. The AI helps you articulate strategy, write evidence-backed research artifacts, explain what changed, and propose reviewable improvements. Strategy activation, candidate promotion, memo selection, and portfolio actions stay yours, made through explicit approvals.
This public repository contains the runnable application, tests, public website, and screenshots. Local planning notes, architecture decision logs, audits, evaluation notes, and operator context files are intentionally kept out of the published tree.
The core loop:
Home Chat ──► Constitution (versioned criteria, deterministic wiring)
│
▼
Runs: Screener ─► Thesis ─► IC Review ─► Investment Memo
│ │
├──── evidence + artifacts┤
▼ ▼
Markets / Company Pages Thesis Health
│ │
└──── Portfolio + Learning ────► Inbox
| Surface | What it does |
|---|---|
| Home | The command surface: daily briefing, strategy chat, archive Q&A, data questions, slash commands, and the always-available conversation drawer. |
| Inbox | The unresolved decision queue: pending Constitution approvals, portfolio pressure, Constitution-fit opportunities, thesis breaks, learning recommendations, and visible operational failures. |
| Runs | Durable stage map for Screener → Thesis → IC Review → Memo. Each stage can run independently, resume, fail visibly, and hand off explicit selections downstream. |
| Markets | Sector, industry, watchlist, and thematic research over retained local data, with bounded cited AI research runs when you ask for deeper work. |
| Screener | Deterministically applies your Constitution's screening requirements to the active universe and ranks survivors by your approved blend. |
| Thesis | One-page AI-written opportunity arguments with explicit return-source decomposition, ranked by return profile for IC Review. |
| IC Review | The memo-worthiness gate: hard hurdles first, then a scored blend of conviction, Constitution fit, and data quality. You can override either way. |
| Memo | Structured seven-section Investment Memos plus a separate machine-checkable monitoring plan. |
| Company Page | Read-only ticker dossier: workflow history map, retained financials, filings/events, ownership evidence, artifacts, and memo-backed thesis health. |
| Library | Ticker-first archive lookup plus retained conversation history; opens the Company Page for any ticker FundOps has retained history on. |
| Portfolio | Ledger-first: purchase lots and sales in, holdings and realized/unrealized P&L projected out. Held positions can queue memo-backed thesis coverage. |
| Settings | Operations only: providers, models, usage records, sync status, data export, and destructive resets. Strategy never lives here. |
Everything generated is a structured, versioned, source-linked artifact tied to the exact Constitution version, evidence bundle, and model steps that produced it.
Prerequisites: Python 3.12+, Node.js 18+
git clone https://github.com/jhchang0407-lang/fundops.git
cd fundops
npm install # sets up Python venv + backend + frontend
npm start # → http://localhost:8000FundOps can run its model work three ways:
- OpenAI API — set
OPENAI_API_KEYin your environment for direct API access. - Your coding agent (Claude Code / Codex) — point FundOps at the coding-agent CLI you already subscribe to, in Settings → Connected services. No API key needed; FundOps invokes the CLI headlessly and only when you explicitly choose it.
- Offline stub — with neither configured, FundOps runs in a deterministic offline mode (clearly marked in provenance) so you can explore the full workflow shape first.
- Describe your strategy in Chat — e.g. "Quality compounders at reasonable prices: ROIC above 15%, gross margin above 40%, low debt. Rank by FCF yield. Only memo-worthy ideas with 15%+ expected return." FundOps drafts a Constitution: exact rules, plain-English interpretations, and a preview of how each workflow gets wired. Nothing activates until you approve.
- Run the workflow — Screener → Thesis → IC Review → Memo, stage by stage or as one pipeline run. Promote/dismiss candidates at every stage; your selections are remembered as learning signals, never as silent rule changes.
- Enter your holdings — purchase lots and sales. FundOps projects P&L from the ledger and queues memo-backed thesis coverage for anything you hold.
- Watch the Inbox — thesis breaks, portfolio pressure, Constitution-fit opportunities, and learning recommendations arrive as reviewable, evidence-first items.
FundOps is bulk-first: breadth data arrives as official bulk products, downloaded once and kept current with tiny daily index ticks. Live APIs are reserved for on-demand research — full filing text for memos, fresh quotes during interactive runs.
| Source | Cost | Role |
|---|---|---|
SEC companyfacts.zip |
Free | Reported fundamentals for the whole universe — one bootstrap download, refreshed weekly or on demand |
| SEC daily index files | Free | ~1–3 MB/day filing detection: who filed drives fact top-ups and thesis-health recalcs for exactly the affected tickers |
| SEC quarterly ownership data sets | Free | Insider transaction evidence for known entities |
| Yahoo Finance (batched) | Free | Price history for the universe, downloaded in batches; daily price updates ride the same sync tick |
| OpenAI / your coding agent | Usage-based or your existing subscription | Thesis/IC/memo writing, strategy interpretation. Tiered models (cheap for extraction, strong for deep work); every call is recorded as an AI Usage Record |
FundOps is built for sometimes-on use: on launch it catches up everything since your last session (index files sync from the last recorded day, then targeted top-ups for exactly the tickers that filed while you were away). To stay current even on days you don't open the app, schedule a headless sync with your OS:
npm run sync # catch-up tick now (bootstraps on first run)
# cron example — weekdays at 6:30pm:
# 30 18 * * 1-5 cd ~/Repos/fundops && npm run syncOwnership evidence comes from the same pipeline: insider transactions (Forms 3/4/5 quarterly data sets) and largest 5%+ holders (Schedule 13D/G filings, parsed from their structured XML). S-3/S-4 registration statements are retained as dilution/M&A events in the filings index.
The one-time bootstrap downloads ~2–3 GB; after that, daily ticks are ~1–3 MB. Total storage footprint is roughly 3–5 GB at Russell 2000 scope (raw bulk cache + workspace database).
backend/
├── core/ # workspace DB + migrations, AI gateway (tiered, recorded, stubbed offline), config
├── domain/ # pure deterministic logic: criteria, guardrails, wiring, IC gate math,
│ # thesis-health evaluation, ledger math, artifact contracts, metric catalog
├── stores/ # platform stores — the ONLY write path to the workspace database
├── services/ # application services: market data, portfolio, inbox projection, strategy
│ └── ingest/ # bulk-first ingestion: SEC companyfacts + daily indexes + ownership, batched prices
├── workflows/ # durable workflow runs: screener, thesis, ic_review, memo, thesis_health,
│ # learning, pipeline
├── chat/ # FundOps Chat: strategy chat + archive Q&A
├── connectors/ # SEC EDGAR + Yahoo Finance adapters
└── api/ # thin FastAPI route adapters + SPA serving
frontend/src/ # React 19 + TS: Home, Inbox, Runs, Markets, workflow pages, Company Page,
# Library, Portfolio, Settings, Artifact Reader — custom app design system
tests/platform/ # backend invariants: ledger math, guardrails, IC scoring, thesis health,
# workflow contracts, chat behaviors
Key invariants:
- One workspace, one owner, one active Constitution, one primary portfolio. Constitution versions are immutable; activation requires an explicitly accepted proposal that passed deterministic guardrails.
- Reported facts ≠ calculated observations. Both are retained with lineage; corrections supersede rather than overwrite; every calculated value records the metric-catalog version.
- Artifacts are append-only. Historical outputs are never edited in place — new versions supersede, and every artifact records its evidence bundle and Constitution version.
- Projections are rebuildable. Holdings, inbox items, library lookups, and latest financials are derived views over retained records, never independent truth.
- Operational failure ≠ investment judgment. Retries and failures stay visible as operational state; they never become verdicts or learning evidence.
The public website is served from docs/index.html; screenshots live in docs/assets/screenshots/.
npm test # backend tests + frontend typecheck
.venv/bin/python -m pytest tests/platform -q # backend platform tests only
cd frontend && npm run dev # frontend dev server with API proxy
cd frontend && npm run build # production build| Variable | Required | Description |
|---|---|---|
OPENAI_API_KEY |
No | Enables AI generation (offline stub mode without it) |
FUNDOPS_DB |
No | Workspace database path (default ~/.fundops/workspace.db) |
FUNDOPS_CONFIG |
No | Operational config path (default ~/.fundops/config.yaml) |
FUNDOPS_CACHE |
No | Bulk data cache directory (default ~/.fundops/cache) |
FUNDOPS_SECRETS |
No | Credentials file path (default ~/.fundops/credentials.yaml) |
FUNDOPS_AI_PROVIDER |
No | Override the AI provider: openai | agent_cli | stub |
SEC_USER_AGENT |
No | User-Agent for SEC EDGAR requests, per SEC fair-use policy |
MIT