KALA

KALA is an investigative web interface for exploring declassified UAP-related archive material. It combines document metadata, DOD video references, geospatial clustering, document relationship graphs, correlation dashboards, AI-assisted document analysis, and date-based resonance mapping into a single React/Vite application.

The project is designed for readers who need to understand an archive before drawing conclusions from it. Instead of presenting files as a flat list, KALA lets a reader ask: where did this record come from, what era does it belong to, what location does it describe, what other records are nearby in the data model, and how confident should an AI-assisted interpretation be?

Research Context

KALA was developed under CreativeDev.Lab's SciFact to SciFi Interfaces research program at CreativeLabTH Group.

The interface uses the language of intelligence displays, star maps, relationship systems, and mission dashboards because the source material is archival, agency-heavy, and pattern-oriented. The visual design is cinematic, but the product goal is research support: make the archive easier to inspect, filter, connect, and question.

What A Reader Can Do

• Start with the hero and archive overview to understand the size and time span of the collection.
• Use agency, era, type, and location summaries to see which parts of the archive dominate.
• Open the geospatial map to find repeated incident zones and location clusters.
• Explore the document network graph to inspect relationships between records.
• Use the correlation dashboard to compare years, agencies, document types, and locations.
• Search the document vault by title, id, location, filename, agency, type, and era.
• Open DOD video metadata and embedded DVIDS streams where public video ids are available.
• Use the birth date resonance portal to find historically significant UAP events near a given date.
• Click any DECODE WITH AI button from the vault, network graph, globe, or resonance portal to auto-trigger KALA AI analysis — no manual re-selection required.

KALA does not try to prove a claim by itself. It is an evidence navigation layer. The application helps the reader decide which records are worth opening, comparing, or questioning next.

Core Principles

• Context first: every archive record should be connected to an agency, era, type, location, filename, size, redaction state, and related records.
• Evidence before speculation: AI output is structured, confidence-scored, and meant to support source review rather than replace it.
• Spatial thinking: maps, graphs, timelines, and correlation views are treated as first-class ways to read the same archive.
• Deterministic metadata: counts, filters, relations, and charts should come from stable manifest data whenever possible.
• Cinematic but usable: starfields, scanlines, HUD panels, and the 3D spaceship support atmosphere without blocking navigation or controls.

Current Archive Scope

The current manifest contains:

Area	Current Scope
Documents	130 metadata records
Videos	24 DOD video metadata records
Agencies represented in documents	6
Location zones counted in stats	20
Manifest location keys	22
Year range	1944 to 2026
Redacted records	4
Computed document relations	about 104 with the current manifest

Document counts by agency:

Agency	Count	Notes
FBI	53	Photo sets, 62-HQ-83894 sections, intelligence records
DOW	44	Mission reports, range fouler debriefs, emails, event records
USAF	13	Project Blue Book and related historical records
NASA	12	Apollo-era photo and transcript material
ARMY	6	Historical intelligence and flying disc records
DOS	2	Diplomatic cable records

The video archive is modeled separately from documents. Videos use agency key VIDEO, but VIDEO is not included in the document agency count because the document and video collections are separate arrays.

Main Product Areas

Archive Intelligence

The statistics panel gives the reader a fast sense of scale. It shows total files, document count, video count, year span, agency distribution, era distribution, top locations, and redaction count.

Global Incident Map

The map groups documents by known geographic location. Unknown and Space are intentionally excluded from map markers. Selecting a marker opens a panel listing related documents for that zone. Clicking DECODE WITH AI from that panel sends the first document from the zone directly into KALA AI.

Document Network

The network graph uses D3 force simulation. Nodes represent documents, node color follows agency color, and edges represent computed relationships from shared location, agency, era, and type metadata. Clicking DECODE on any selected node sends that document directly into KALA AI.

Relations are generated from shared metadata:

same known location -> +3.0
same agency         -> +1.0
same era            -> +1.0
same type           -> +0.5
emit edge if weight >= 3

Data Correlation

The correlation dashboard gives the reader multiple chart views over the same manifest:

• incidents by year
• agency distribution
• top incident zones
• document type by era
• top agency, top location, peak year, and redaction count

These charts are not a statistical model. They are exploratory views that help a reader notice concentrations and gaps.

Document Vault

The vault is the main document browsing surface. It supports search, agency filter, era filter, pagination, document details, and file metadata.

Selecting a document opens a modal with:

• document id
• agency
• type
• year
• location
• era
• file size
• redaction status

The DECODE WITH AI action fires the API call immediately via startDecode and scrolls to the decoder section. Analysis begins without any manual re-selection in the decoder.

UAP Footage

The video archive displays 24 DOD video metadata records. Some records have known DVIDS video ids and can be embedded or opened on DVIDS. Records without known public video ids fall back to a DVIDS search link.

KALA AI Decoder

The decoder supports four analysis modes:

• metadata-only assessment from a selected manifest record
• text analysis for .txt and .md
• image analysis for common image formats
• video frame analysis by extracting representative frames client-side

PDF handling currently sends a binary/base64-derived placeholder rather than true PDF text extraction.

KALA AI returns structured JSON with:

• executive summary
• visual description
• classification era
• document type
• key entities
• incident details
• significance
• significance reason
• related topics
• redaction level
• confidence score

Metadata-only analysis is required to keep confidence below 0.4 because no document body, image evidence, or video frame evidence has been decoded.

When a decode is triggered from any section, a toast notification appears at the top of the page showing the document title and a progress indicator while the request is in flight.

Birth Date Resonance Portal

The resonance portal lets a reader enter a birth date and discover historically significant UAP events close to that date. It shows:

• zodiac sign and UAP archetype profile
• nearest sightings by calendar proximity
• related document records
• direct decode handoff into KALA AI

Tech Stack

• React 18
• Vite 5
• Tailwind CSS 4 through @tailwindcss/vite
• Three.js, @react-three/fiber, and @react-three/drei
• D3 for the document network
• Leaflet and React Leaflet for the map
• Framer Motion for transitions
• Anthropic SDK behind the serverless KALA AI API routes

Getting Started

Install dependencies:

npm install

Run the Vite dev server:

npm run dev

Build for production:

npm run build

Preview the production build:

npm run preview

Environment

Create .env from .env.example.

Required for AI analysis routes:

ANTHROPIC_API_KEY=your_api_key_here

The user-facing product name is KALA AI. The current backend provider is configured in the API routes and should stay out of visible UI copy.

API Routes

api/analyze.js

Handles structured analysis for selected metadata, text content, images, and sampled video frames. The route expects a POST request and returns { success, analysis, docId } when successful.

api/meta-analyze.js

Handles collection-level questions over manifest metadata. It imports the manifest server-side, limits the prompt to the first 50 matching rows, asks the model to answer with document ids, then strips basic markdown characters from the returned text.

Project Structure

api/
  analyze.js          Structured KALA AI analysis endpoint
  meta-analyze.js     Metadata question-answer endpoint
public/
  spaceship.gltf      3D spaceship model used in the hero section
src/
  App.jsx             App shell, boot loader, section layout, lazy loading, DecodeToast
  components/         UI sections, visual systems, archive tools
  data/
    manifest.js       Canonical archive metadata and relation functions
    decodeStore.js    Module-level decode state; fires API calls on button click
    sightings.js      UAP sighting records and zodiac/date utilities
  index.css           Global styling, fonts, layout, HUD visual system
vite.config.js        Vite, Tailwind plugin, dev proxy, chunk splitting
vercel.json           Deployment config for Vercel serverless API routes

Important Implementation Details

The app is a single-page, section-based interface. Section ids are defined in App.jsx and used as scroll targets:

hero
stats
globe
network
correlation
vault
video
decoder
birthdate

The global starfield is mounted once behind all content. The 3D spaceship is mounted only in the hero section. Decorative canvas layers should not capture pointer events.

The manifest is the source of truth for records, agencies, types, eras, locations, videos, computed relations, computed communities, and summary stats.

Decode Handoff Pattern

src/data/decodeStore.js coordinates the decode trigger across all sections without requiring the components to know about each other.

When any DECODE WITH AI button is clicked:

1. startDecode(doc) fires the /api/analyze POST immediately.
2. A kala:decode-start custom DOM event is dispatched — DecodeToast in App.jsx listens to this and shows a toast notification.
3. If AIDecoder is already mounted, the promise and doc are pushed directly via a registered handler.
4. If AIDecoder is not yet mounted, the doc and promise are stashed and consumed when it mounts.

This means the API call is already in flight by the time the scroll animation ends.

Data Model Summary

Documents are stored in DOCUMENTS with this shape:

{
  id: "DOW-D3",
  agency: "DOW",
  type: "mission-report",
  year: 2020,
  title: "Mission Report - Arabian Gulf",
  location: "Arabian Gulf",
  filename: "DOW-UAP-D3-Mission-Report-Arabian-Gulf-2020.pdf",
  size: 101111,
  redacted: false
}

The era field is derived automatically from year.

Videos are stored in VIDEOS with this shape:

{
  id: "DOD-V01",
  dvidshubId: "111688723",
  title: "UAP Encounter - Arabian Gulf",
  location: "Arabian Gulf",
  year: 2020,
  agency: "VIDEO",
  size: 6814452
}

Known Caveats

• Unknown years currently classify as WWII because JavaScript evaluates null <= 1945 as true.
• Some source strings contain mojibake from encoding issues. Clean these carefully instead of running broad replacements.
• FBI B-series photo sizes use runtime randomness in the manifest. That makes totalSizeGB non-deterministic between runs.
• PDF upload support does not currently extract semantic PDF text.
• The Vite dev proxy points /api to http://localhost:3001, but Vercel-style API routes are not automatically served by plain Vite. Local API testing may require a compatible serverless/dev setup.
• Some video records have DVIDS ids in manifest metadata and a separate embed id map in VideoArchive.jsx; keep those aligned when adding videos.

Development Guidance

• Keep user-facing AI language as KALA AI.
• Keep provider names in backend or developer documentation, not visible product copy.
• Add archive records in src/data/manifest.js with stable ids and deterministic sizes.
• Add new locations to LOCATIONS before assigning documents to them.
• Use existing agency and type keys where possible.
• Run npm run build after documentation or code changes that could reveal import, syntax, or bundling issues.

License

See LICENSE.