How Avichive works
Avichive began with decades of files scattered across old drives, backups, and project folders. Finding a file was difficult. Knowing which copies were identical, which formats still opened, or which projects could be recovered was harder.
The first step was an inventory. A Go scanner recorded file paths, sizes, content hashes, and available metadata in PostgreSQL. A Rails application made that inventory browsable. That foundation now supports cloud-preserved originals, AI-assisted curation, overlapping life eras, photo and people review, playable media, and restored websites.
This guide explains the system behind those features: what each pipeline reads, what it creates, where human decisions enter, and how the archive preserves the evidence those decisions depend on.
One archive, several ways to understand it
The same photograph might have three historical paths, one content identity, a preserved cloud original, a display image, two era memberships, and several reviewed faces. Those records answer different questions. Where was it found? Which bytes should be preserved? What part of life does it belong to? Who appears in it?
Avichive keeps those questions connected without collapsing them into one category or one folder.
flowchart TD
A[Files on historical drives] --> B[Scanner inventory]
B -->|Recorded hash| C[Unique content and occurrences]
C --> D[Preserved cloud originals]
C --> E[Browse and search]
D --> F[Viewable photos and playable media]
C --> G[File analysis and curation]
C --> H[Life-era context]
F --> I[Face detection and people review]
D --> J[Recovered source and assets]
J --> K[Restored historical projects]
Preservation and interpretation can proceed independently. A keep/skip decision is recorded curation; it does not itself delete bytes. An era suggestion is evidence for review; it does not automatically assign membership. A face match is a candidate; a person’s identity is confirmed by the owner.
Explore the system
| Start with | What it explains |
|---|---|
| How it works | Scanner ownership, content identity, storage, analysis, and data flow |
| Scanning the archive | The original discovery pipeline and the facts it records |
| Browsing and search | Directory and media views, duplicates, and approximate filename lookup |
| File analysis and curation | Rules, bounded content samples, structured AI results, and keep/skip review |
| Life eras | Overlapping biographical context and explicit inclusion/exclusion |
| Photo analysis and people | Local face detection, manual identification, and reviewed match candidates |
| Cloud preservation | Canonical originals, display derivatives, and catalog backups |
For the original motivation, read the archive’s history. For engineering details, follow the foundations through scanning, deduplication, and the data model.
Preservation makes reconstruction possible
Some files become useful again through more than a download. Recovered databases, writing, images, and code can be transformed into a working historical experience.
- flombaum.com turns recovered early-2000s writing and media into readable sections served by Rails.
- Tweefle recreates a 2009 Twitter wall with archival data rather than a live social feed.
- Sawbaum reconstructs a Vine montage interface while explaining the missing original media and replacement footage.
The restoration guide traces what survived, what changed representation, and what had to be reconstructed.
Scale, with a date attached
The July 22, 2026 inventory documented 3,576,948 scanner records and 1,185,610 unique contents. The July 24 upload checkpoint documented 1,185,605 uploaded originals and five intentionally skipped contents. These are historical snapshots, not live counters. Generated previews, analyzed contents, and reviewed people have separate coverage.
See capability status and evidence for the dated inventory and the distinction between implemented features, recorded outcomes, and ongoing research. Semantic photo search and captions remain research rather than features implied by the presence of embeddings or AI infrastructure.