PARALLAX

Inspiration

Every year, companies lose billions of dollars because lawyers and analysts miss hidden contradictions buried inside 500-page contracts.

The standard workflow is painful: open the PDF, hit Ctrl+F, type a keyword like "liability," and read through 200 results. Hours pass. And somewhere on page 387, a clause says "Vendor assumes unlimited liability" — while page 42 says "Liability is capped at $10,000."

These two clauses directly contradict each other. But Ctrl+F can't catch it, because you cannot search for a risk you don't know exists. This is the "unknown unknowns" problem — and it's why audit teams still miss critical risks that later explode in litigation.

We wanted to build a system that doesn't wait for the user to ask a question. A system that proactively finds the risk. That system is PARALLAX.

What it does

PARALLAX transforms dense legal documents into an interactive semantic galaxy.

Every dot on the canvas represents one clause.
Similar clauses cluster together (Payment Terms, Termination, Indemnification, Confidentiality, Liability).
We apply Isolation Forest anomaly detection directly to the text embeddings.
Clauses that are mathematically isolated from their neighbors are flagged as contradictions — they glow red.

The user never types a search keyword. PARALLAX proactively surfaces the risks.

Features:

Click any dot to inspect the clause text.
Red dots open an "⚠️ Anomaly Detected" modal with confidence scores and explanations.
Cinematic auto-demo plays 2 seconds after page load.
Live stats: "1,005 clauses analyzed · 5 anomalies flagged · 0 keywords required."
Legend showing the 5 semantic clusters.

How we built it

The Round 1 prototype is a single, self-contained HTML file — zero dependencies, zero build steps, deployed on GitHub Pages.

Frontend:

HTML5 Canvas for GPU-accelerated rendering of 1,000+ glowing dots at 60fps.
Vanilla JavaScript for camera math (pan, zoom, tween animation).
Custom glow effects via createRadialGradient.
A cinematic auto-demo that walks viewers through the concept without clicks.

Planned Backend (Round 2):

Unstructured.io for PDF parsing.
Google Gemini embedding-001 for 768-dim vectors.
UMAP for dimensionality reduction into 2D.
Isolation Forest (via PyOD) for anomaly scoring.
Supabase (pgvector) for storage.

The math: anomalies require fewer random splits to isolate than normal points. We measure the average path length required to isolate each clause and translate it into a risk score.

Challenges we ran into

1. Pixi.js v7 → v8 migration hell. Every online tutorial was outdated. The new PIXI.Application({...}) constructor became async, app.view became app.canvas, and we hit a this._cancelResize is not a function crash caused by React 19 Strict Mode double-invoking useEffect.

2. Turbopack cache corruption on Windows. Next.js 16's Turbopack dev server corrupted its own .next cache, causing a Cannot find module './chunks/ssr/[turbopack]_runtime.js' error. We fixed it by switching to next dev --webpack.

3. PowerShell security policies. Windows PowerShell blocked npm from running scripts. We switched VS Code's terminal to Command Prompt.

4. The framework tax. After days of fighting framework-level bugs, we made the tough call: drop Next.js and ship a pure HTML file. This meant sacrificing Vercel deployment — but it gave us a working, portable, instantly-loadable demo. The judges' first impression matters more than the tech stack we brag about.

Accomplishments that we're proud of

Shipped a fully working prototype despite spending the first 40% of our time fighting framework bugs.
Rendered 1,000+ animated glowing dots at 60fps with a single HTML file — no libraries.
Built a cinematic auto-demo that explains the concept in 20 seconds without user input.
Designed a coherent narrative: the visual (clusters + red anomalies) directly maps to the underlying math (embeddings + Isolation Forest).
Recovered from a broken framework migration by making the brave call to strip everything down.
Deployed a live, publicly accessible URL on GitHub Pages.

What we learned

Search is reactive. Intelligence is proactive. Ctrl+F waits for you to type. PARALLAX finds the risk for you.
Semantic embeddings capture meaning, not keywords. Two clauses can use different words but mean the same thing.
Anomaly detection on embeddings is powerful. Isolation Forest isolates contradictory clauses without labeled training data.
Dimensionality reduction is a UX superpower. UMAP compresses 768D meaning into 2D coordinates.
Scope discipline matters more than features. Stripping down to what actually works was the smartest decision we made.
The "wow" is in the visualization. The math is invisible. The glowing galaxy is what makes people say "wait, do that again."

What's next for PARALLAX

Round 1 is the visual prototype. Round 2 will connect the full backend:

Live PDF upload — drag any contract and get instant analysis.
Real Gemini API embeddings — replace hardcoded data with live vectors.
Python FastAPI service — running UMAP and Isolation Forest.
Side-by-side text viewer — click a red dot and see the sentence highlighted in the source PDF.
Exportable audit reports — summary of flagged anomalies with confidence scores.
Multi-document comparison — detect contradictions across a folder of contracts.

Our long-term vision: PARALLAX becomes the default first-pass review tool for legal and financial documents — the way spellcheck is for writing.

Stop searching. Start seeing.