I built a open source free stock analysis tool with 55+ data dimensions

#opensource #ai #javascript #showdev

The Problem

I invest in stocks on the side. My workflow looked like this:

Open Screener.in for fundamentals
Open TradingView for technicals
Ask Perplexity for a quick summary
Manually check insider trading on SEC EDGAR
Look up institutional holders somewhere else

Five tools.
None of them talked to each other.
And I was still missing actual valuation models, moat analysis, and any kind of fact-checking on the AI summaries.

Bloomberg does all of this in one place. It also costs $24,000 per year.

So I built one tool that does it for free.

What It Does

You enter a stock ticker. Eight seconds later, you get a report with:

Nipun Score - A proprietary A+ to F letter grade based on a weighted composite of technicals (25%), fundamentals (25%), sentiment (20%), risk (15%), and insider activity (15%)
Three valuation models - DCF, Benjamin Graham Number, Peter Lynch Fair Value
Scenario analysis - Bull, Base, and Bear price targets with probability estimates
Competitive moat - Wide, Narrow, or None with moat sources (brand, network effects, switching costs, etc.)
SWOT analysis - AI-generated strengths, weaknesses, opportunities, threats
Insider trading - Recent executive buys and sells with dollar amounts
Institutional ownership - Top 10 holders with position changes
...and about 20 more sections

The Architecture

The backend is a serverless worker that runs a 4-phase pipeline:

Phase 1 - Data Collection (~2-3s)
10+ parallel API calls to Finnhub (financials, peers, insider trades, earnings), SEC EDGAR (10-K, 10-Q filings), Reddit RSS (sentiment), and Yahoo RSS (news).

Phase 2 - Compute (~5ms)
Zero API calls.
Pure math.
This phase calculates Altman Z-Score, Piotroski F-Score, DCF valuation, Graham Number, momentum scoring, risk-reward ratios, and dividend safety. All deterministic.

Phase 3 - AI Synthesis (~3-5s)
Two parallel calls to Google Gemini.
One generates the main analysis.
The other generates premium insights (scenario analysis, moat, SWOT, investment thesis). I built a 5-model cascade: 2.5 Pro → 2.0 Flash → 1.5 Pro → Flash → Lite. If one model rate-limits, it falls through to the next automatically.

Phase 4 - Second Opinions (~1-2s)
Cerebras generates a contrarian take.
Cohere runs a fact audit, classifying every AI-generated claim as grounded, speculative, or unverifiable. Both are non-fatal — if they fail, the report still ships.

Security

This was important to me. The tool uses a BYOK (Bring Your Own Keys) model:

API keys encrypted client-side with AES-256-GCM
Key derivation: PBKDF2 with SHA-256, 100K iterations
16-byte random salt, 12-byte random IV
Keys sent via X-Nipun-Keys header, never in request body
Worker processes keys in memory, never persists them
Everything uses the Web Crypto API — zero npm dependencies for crypto

The Interesting Technical Decisions

1. Mock data fallback on every API call

Every single external call is wrapped in try/catch with a fallback to mock data. This means:

The app literally cannot crash from a bad API response
Demo mode works using the same fallback path
You can run the full UI without any API keys

2. Cascading AI models

Instead of relying on one model, I chain five. Rate limits, outages, and model-specific failures are handled transparently. The user never sees an error.

3. Contrarian AI

Most AI tools agree with themselves. I deliberately ask a second model to disagree with the first. This gives you both sides of the argument instead of just confirmation bias.