AfterTrace 如何結合 Gemma 4 挖掘被遺忘的網絡歷史

What if we could travel back through the internet's history and discover what happened to the websites and software projects we left behind?

The internet is constantly evolving. Websites disappear, communities become inactive, and open-source repositories stop receiving updates. But their digital footprints often remain scattered across archived pages, repository histories, and forgotten documentation.

I wanted to build something that could explore these forgotten traces.

That idea became AfterTrace — The Digital Archaeology Lab.

What I Built

AfterTrace is an AI-powered digital archaeology platform that investigates the historical evolution of websites and open-source software.

It uses real historical data from the Internet Archive and GitHub, combined with Google's open-weight Gemma 4 model, to reconstruct digital histories and generate evidence-backed research reports.

The application has two main modes.

Website Archaeology

Enter a public website URL, and AfterTrace retrieves available historical captures from the Internet Archive Wayback Machine.

Features include:

Historical website snapshots and timestamps.

Extracted content from archived pages.

Historical evolution timelines.

Comparisons between two historical snapshots.

AI-generated historical narratives.

GitHub Archaeology

Enter a public GitHub repository URL to investigate its development history.

AfterTrace retrieves:

Repository metadata and statistics.

README documentation.

Recent commit history.

Development and maintenance signals.

Potential preservation and modernization opportunities.

How I Interpreted Digital Archaeology

I see forgotten websites and inactive repositories as digital memories.

A website created in the 1990s might contain traces of an online community that no longer exists. An old repository might preserve engineering decisions that are still useful today.

AfterTrace attempts to bring these fragments together.

The goal is not simply to retrieve old data. It is to make that data understandable, searchable, comparable, and useful.

The Moment It Came Together

One of the most interesting moments during development was testing AfterTrace against GeoCities.

The application successfully retrieved historical captures from:

October 22, 1996

July 3, 1998

January 25, 1999

Seeing actual historical website content from different years was fascinating.

The Snapshot Comparator also revealed measurable textual differences between the 1996 and 1999 captures:

37 lines added.

5 lines removed.

It turned a forgotten website address into something we could actually investigate.

Gemma 4: The Intelligence Engine

Gemma 4 is a central part of AfterTrace.

I integrated Google's official @google/genai SDK to use the gemma-4-26b-a4b-it model for historical analysis.

Gemma receives bounded historical evidence and generates structured interpretations, including:

Historical identity and evolution.

Notable changes and decay signals.

Repository development insights.

Preservation recommendations.

Digital Resurrection Blueprints.

However, AI-generated information should not automatically be treated as historical fact.

That brings me to one of the project's most important engineering components.

Deterministic Evidence Ledger

One of my main objectives was to make AfterTrace evidence-driven rather than just another AI chatbot.

I implemented a Deterministic Evidence Ledger that validates AI-generated claims against actual source records.

The system separates information into three categories:

VERIFIED: Claims supported by available primary evidence.

AI_INFERRED: Reasonable interpretations generated by the model but not independently established as facts.

UNKNOWN: Information that cannot be established from the available evidence.

If a generated claim references an invalid or nonexistent source, the application downgrades it instead of presenting it as verified.

My core engineering principle is:

AI interprets evidence; deterministic code establishes facts.

Historical Snapshot Comparator

AfterTrace also includes an interactive comparison engine.

Users can select two historical website captures and compare their extracted textual content.

The system identifies:

Added content.

Removed content.

Changes between historical versions.

The comparison works on extracted textual data without executing archived website scripts.

Why Open Innovation Matters

I believe digital preservation should be accessible to developers, researchers, students, and independent communities.

Open innovation makes it possible to build on publicly accessible archives, inspect source code, experiment with open-weight AI models, and improve existing tools collaboratively.

AfterTrace combines:

The Internet Archive's historical web data.

GitHub's public repository information.

Gemma 4's open-weight AI capabilities.

An independently developed evidence-validation system.

The combination demonstrates how existing open technologies can be integrated to address a real problem.

I also want other developers to be able to explore the source code, understand the architecture, and contribute improvements.

Technology Stack

Component

Technology

Frontend

Next.js 15

Language

TypeScript

Styling

Tailwind CSS

AI Model

Google Gemma 4

AI SDK

@google/genai

Historical Data

Internet Archive Wayback CDX API

Repository Data

GitHub REST API

HTML Extraction

Cheerio

Validation

Zod

Testing

Vitest

Deployment

Vercel

Security and Engineering Challenges

Building AfterTrace involved several engineering challenges.

SSRF Protection

Since users can submit website URLs, I implemented URL validation and server-side protections against requests to private networks, localhost, and cloud metadata endpoints.

Structured AI Responses

Gemma responses are validated against structured schemas. Malformed responses are handled rather than silently accepted.

Evidence Validation

The Evidence Ledger ensures that AI-generated claims cannot introduce fabricated source identifiers and file references.

API Rate Limits

GitHub API requests are bounded and handled with appropriate error states.

Inference Latency

Large historical HTML payloads can make AI analysis slow. I introduced bounded evidence extraction and prompt optimization to reduce unnecessary input.

Testing and Verification

After implementing the application, I ran an automated and live verification process.

Test

Result

Automated test suite

30/30 passed

TypeScript typecheck

0 errors

Production build

Passed

Live Wayback integration

Passed

Live GitHub integration

Passed

Gemma 4 inference

Passed

SSRF security tests

Passed

Snapshot comparison

Passed

The application is deployed and publicly accessible.

What's Next?

AfterTrace is an evolving project. Some areas I would like to explore further include:

More historical archive providers.

Better website reconstruction.

Deeper repository evolution analysis.

Improved preservation workflows.

More advanced historical visualization.

Explore AfterTrace

Live Application:
https://after-trace-delta.vercel.app

GitHub Repository:
https://github.com/rudraarcadeskills-art/AfterTrace

I built AfterTrace with one simple idea in mind:

The internet has a memory. Even when websites disappear, their stories deserve to be explored.

Thanks for reading!