I Built a Browser-Based Python Quant Research Sandbox on Cloudflare's Free Tier

#python #cloudflare #opensource #fintech

"Can you build a Python quant research environment that runs entirely in the browser, on Cloudflare's Free Tier alone?"

That question cost me several nights of sleep. The answer, it turns out, is yes — and today I'm open-sourcing the result.

VibeQuant Browser is a browser-based quant research sandbox inspired by the workflow of GS Quant. You type a trading idea in natural language, an LLM turns it into runnable Python, and the code executes right in your browser against real market data — no backend Python server, no signup, no cost.

🔗 Live demo: https://vibequant-web.pages.dev/
📖 User manual: USER_MANUAL.md
⭐ GitHub: gameworkerkim/vibe-investing → VibeQuant

One thing before we start: this is **not* a GS Quant replacement, and it's research/education only — not investment advice. It's an independent open-source project (Apache 2.0), not affiliated with Goldman Sachs.*

What it does

LLM → Quant → Python. Describe a strategy in plain language; the LLM (DeepSeek, gated to finance-only prompts) generates runnable vi_browser Python and executes it immediately.
Python in the browser. Pyodide (CPython compiled to WebAssembly) runs your scripts client-side. Your machine is the compute layer.
Real market data API. Candles and quotes served by a Cloudflare Worker (Yahoo Finance provider), cached in R2 with D1 indexes.
Interactive charts for equity curves, indicators, and price series.
Technical indicators out of the box: MA, RSI, MACD, Bollinger Bands, momentum, volatility, max drawdown — plus an educational backtest().
100% Cloudflare Free Tier: Pages · Workers · D1 · R2 · CDN & Cache API. Monthly bill: $0.

The workspace: prompt on the left, Python on the right

The workspace splits into an LLM prompt pane and a Python editor, with Result / Error log / Chart panes below. The LLM round-trip works like this:

Your prompt hits DeepSeek through a Cloudflare Worker (API keys never touch the browser)
If Python comes back, it auto-fills the editor
It runs immediately in Pyodide — prints go to Result, show_chart(...) renders to Chart, exceptions land in the Error log

After that, you iterate on the code directly — no need to call the LLM again for every tweak.

Try a prompt like this in the live demo:

Compare 22-day momentum for NVDA, MU, SNDK, AVGO.
Rank only computable names; exclude N/A.
Build vi_browser Python and chart series with show_chart.

Or write the Python yourself:

# Runs in Pyodide — in your browser, not on a server
from vi_browser import get_candles, ma_cross_signal, backtest, show_chart

candles = await get_candles("005930.KS", days=180)   # → Cloudflare Worker API
signals = ma_cross_signal(candles, fast=10, slow=30)
result = backtest(candles, signals, fee_bps=10)

print(result["metrics"])
show_chart(result["equity"], title="Equity", series_label="equity")

There's also an Examples section with one-click sample chips (momentum, RSI zones, MA cross, multifactor) and a full GS Quant ↔ VI Quant API mapping table, so you can see exactly which familiar APIs run in the browser, which run locally via pip, and which are still planned:

The architecture: free tier as a design constraint

The core design decision: split the product along what the free tier can actually do.

Concern	Where it runs	Why
Market data (quotes, candles, assets)	Cloudflare Workers + Pages + D1 + R2 + CDN	Single platform, free-tier first
Quant computation (scripts, indicators, backtests)	Browser Python via Pyodide (WASM)	Workers can't run full Python; no server-side `exec`

Cloudflare Workers give you 100k requests/day but only 10 ms of CPU per invocation on the free plan. That kills any dream of running pandas server-side — so I didn't. The Worker does one thing: serve candles fast. Candle objects live in R2, D1 holds only indexes, and the Cache API absorbs hot paths. All the actual number-crunching moves to the user's browser via WebAssembly.

This constraint turned out to be a feature: there is no server that executes user code, which makes the security story dramatically simpler (SECURITY.md).

What honestly doesn't work (yet)

I'd rather you find this here than in production:

Claim you might assume	Reality
Full GS Quant API compatibility	A subset of public APIs; many symbols are stubs today
Same numbers as GS / Marquee	Never promised — different data, different models
Full `vi_quant` in the browser	No — only a thin WASM-safe subset (`vi_browser`)
Derivative pricing (QuantLib)	Not started
Unlimited free-tier data ingest	No — see LIMITATIONS.md

Status: Pre-Alpha. The committee demo stage is usable; it is not a production research engine.

Why I built this

Retail investors in Korea — my home market — are famously aggressive with leverage. I wanted to lower the barrier to quantitative thinking: give anyone a free, open place to test an idea against data before betting real money on a hunch.

But the longer-term goal is bigger. Today VibeQuant is a single-player sandbox. The roadmap points toward a Multi-LLM Quant Committee: multiple LLMs receiving the same market data, generating investment hypotheses, reproducing them as Python, cross-verifying each other's results, and evaluating Alpha alongside Risk (VaR). Same APIs, same data, reproducible verification — that's why the GS Quant-style API surface matters.

Still ahead: a proper backtest engine, factor research, and community-driven strategy sharing and verification. The first foundation, though, is done.