Zero-origin. Sovereign AI. No kill switch.

The Architecture in One Sentence

80+ Cloudflare Workers serve 47 languages from 330+ edge locations, backed by 49+ D1 databases, 20+ R2 storage buckets, and a local NVIDIA RTX 5090 generating sovereign voice and video content. The platform has no origin server, no containers, no load balancers, and no third-party AI dependencies in the production content pipeline.

Edge Compute: 80+ Cloudflare Workers

Every service in the stack is a Cloudflare Worker — a V8 isolate running at the edge in 330+ cities. Sub-50ms cold starts globally. No containers, no load balancers, no scaling decisions. Each Worker handles its own routing, data access, and response generation.

Cluster 1: Word Orb (The Data Engine)

The foundational vocabulary and knowledge API that powers everything else. Think of it as the brain.

• word-orb-api: Core API serving 162,253 words, 601,143 translations, and 30,288 knowledge graph links
• word-orb-mcp: Model Context Protocol server exposing 12 tools to AI agents (Claude, GPT, LangChain, etc.)
• word-generator: Content generation pipeline for new word entries
• translation-scheduler: Automated translation pipeline across 47 languages
• wordorb-slack: Slack integration delivering daily lessons in Slack channels
• wordorb-telegram: Telegram bot for instant word lookups and daily lessons
• wordorb-education-app: Education-focused API variant for school deployments

Cluster 2: Kelly & The Daily Lesson (The Consumer Product)

An AI tutor named Kelly who delivers personalized daily vocabulary lessons. Think of it as the face.

• kelly-teaches: Primary consumer site at thedailylesson.com — lessons, assessments, enterprise demo pages
• kelly-chat: Real-time conversational AI via Claude Haiku with SSE streaming. Routes: chat.thedailylesson.com, api.kellyai.com
• kelly-audio-prep: Nightly cron pre-generating 150 audio files/day across 10 languages and 3 age groups. Routes: audio.thedailylesson.com
• kelly-shield: 8-layer resilience system from edge cache to satellite delivery to physical books. Routes: shield.kellyai.com
• kelly-video-factory: Multi-engine video orchestrator — Kling 3.0 cinematic, HeyGen batch, LTX-2.3 free local, sovereign pipeline
• kelly-engine: Lesson generation engine — builds personalized content from the Word Orb data layer
• kelly-assets: R2 CDN serving 15,542 media files at kelly-assets.nicoletterankin.workers.dev
• kelly-tts: Text-to-speech endpoint for real-time audio generation
• kellyai-site: Kelly’s brand site at kellyai.com with live /chat interface
• curiouskelly-site: Marketing and subscription site at curiouskelly.com with Stripe integration

Cluster 3: Orwell Observatory (The Mission Project)

Open-source anti-Orwellian global defense infrastructure. 15 deployed APIs covering 195 countries. CC-BY-4.0 licensed.

• orwell-api (EPIC-001): Threat taxonomy — 8 categories, 131 indicators, 55 historical parallels
• orwell-map (EPIC-002): Global risk mapping — 195 countries, Freedom House/RSF/V-Dem scores, 583 index scores, 30 incidents
• orwell-mechanisms (EPIC-003): Control mechanism analysis — 21 mechanisms, 21 counter-measures, 15 surveillance vendors
• orwell-money (EPIC-004): Follow-the-money — 50 entities, 75 contracts ($8.26B), 30 financial flows
• orwell-compliance (EPIC-005): Psychological compliance — 7 mechanisms, 25 behavioral indicators, 18 interventions
• orwell-doublespeak (EPIC-006): Propaganda detection — 223 patterns across 8 categories with real-time POST /analyze
• orwell-audit (EPIC-007): Domestic practices audit — 40 questions, 10 country audits, 400 scores, rankings
• orwell-registry (EPIC-008): Living threat registry — 500 entries across ~60 countries, 690 updates, RSS feed
• orwell-funding (EPIC-009): Journalism funding — 200 applications, 60 grantees ($6.86M), weighted scoring
• orwell-whistleblower (EPIC-010): Whistleblower pipeline — 50 protection laws, 120 attorneys, 100 OPSEC courses in 15 languages
• orwell-archive (EPIC-011): Document archive — R2 storage with SHA-256 hash verification, deletion event tracking
• orwell-provenance (EPIC-012): Media provenance — 80 trust anchors, 300 verifications, deepfake probability scoring
• orwell-courage (EPIC-042): Courage recognition — 250 nominations, 60 awardees, 30 oral histories with consent-level redaction
• orwell-exposure (EPIC-043): Doublespeak exposure — 100 daily highlights, 10 awards, 20 analyses
• orwell-memory (EPIC-044): Intergenerational memory — 50 testimonies, 10 collections, 5 education modules
• orwell-dashboard: Visual monitoring dashboard for all 15 epics
• censorship-map: Interactive d3-geo choropleth of global censorship risk across 195 countries
• orwell-mcp: MCP server exposing all Orwell data to AI agents

Cluster 4: Autonomous Agent Fleet

Five Worker-based agents with D1 database brains, cron schedules, and autonomous Gmail draft generation.

• softy-agent: SoftBank Robotics deal management — 7 skills (watchdog/6h, intelligence sweep/daily, reply handler, follow-up drafter, demo monitor/4h, Stripe sentinel/2h, weekly digest)
• anty-agent: Anthropic outreach — weekly sweep, 6 contacts tracked, IPO timing intelligence, drip sequences
• aly-agent: Cloudflare relationship intelligence — compounding story engine, blog draft management, signal scanning at ada5am.com/aly
• granty-agent: Grant portfolio manager — 14 grants tracked ($110.7M pipeline), MIT Solve countdown, deadline escalation
• vantage-agent: Strategic intelligence — 4 layers (Beacon/Compass/Lens/Helm), anti-pattern detection, decision outcome learning loop at ada5am.com/vantage

Cluster 5: Email & Infrastructure

• email-engine: 6 verified sending domains, 7 HTML templates, 3 sequences, SVIX webhook verification, batch endpoints (100 contacts, 500 sends), domain-aware persona routing
• lotd-healthcheck: 15 endpoints monitored every 5 minutes (6 critical, 5 product, 4 infra) with email alerts on consecutive failures
• wren-*: 5 Workers for crisis communications targeting internet blackout scenarios (Iran, 88M people)
• lotdpbc-site: This corporate site you’re reading — an SSR Worker serving all pages inline

Data: 49+ D1 Databases

Cloudflare D1 is distributed SQLite at the edge. Every query runs at the nearest Cloudflare data center. No connection pooling, no replica lag, no database server to manage.

The Main Database: dailylesson-db-dev

492 MB across 77 tables. This is where all content lives.

• words: 162,253 entries with definitions, etymology, IPA pronunciation, difficulty levels, usage examples
• word_translations: 601,143 translations across 47 languages
• word_ethics: 161,763 ethical content ratings for age-appropriateness and cultural sensitivity
• knowledge_graph: 30,288 weighted links (synonyms, antonyms, related concepts, semantic clusters)
• kelly_lessons: 74,008 base lessons across 19 languages, 4 tracks, 5 phases each
• kelly_lesson_variants: 164,250 archetype-personalized variants (English)
• kelly_interactions: 24,989 assessment items (true/false, multiple choice, open-ended, A/B)
• kelly_looks_v2: 36 visual looks with archetype mappings
• kelly_media: 15,542 media file records (videos, images, audio)
• word_tones: 1,992 tonal variations (55 words × 12 archetypes × 3 ages)
• api_keys: Developer API key registry synced from the email engine signup flow

Purpose-Built Databases

• 14 Orwell epic databases: Each epic gets its own database — isolation prevents a bug in one system from affecting others
• softy-deal-brain: SoftBank deal state machine (deals, interactions, intelligence, drafts tables)
• vantage-intelligence-db: Signals, reports, decisions, anti-patterns, subscribers
• aly-intelligence-db: Relationship timeline, stack metrics, briefings
• email-engine: Queue, logs, templates, contacts, sequences, open/click tracking, bounce management
• kelly-video-factory: Video production queue with multi-engine tracking (engine, engine_id columns)
• lotd_observability: Page view events, beacon data, operator alerts

Media: 20+ R2 Buckets

R2 provides S3-compatible object storage with zero egress fees — a critical advantage when serving 2,300+ videos globally. All media is served through CDN Workers with proper cache headers and CORS.

• kelly-assets: 15,542 files — 2,300+ lesson videos, 36 visual looks across 10 archetypes and 3 ages, 850 legacy video clips, 5 email avatar states × 3 sizes, canonical stills, voice references. Served at kelly-assets.nicoletterankin.workers.dev
• orwell-archive-storage: Document preservation with SHA-256 hash verification for EPIC-011
• orwell-memory-storage: Intergenerational testimony recordings for EPIC-044
• orwell-whistleblower-storage: Secure OPSEC course materials for EPIC-010
• orwell-provenance-storage: Verification evidence chain for EPIC-012
• orwell-courage-storage: Oral history recordings with consent-level access for EPIC-042
• Wren crisis assets: Offline-first content packages for internet blackout scenarios
• Unity CDN: WebGL assets for 3D Kelly avatar experiments (oldest Worker, December 2025)

KV & Cache Layer

Cloudflare KV provides globally-replicated key-value storage with eventual consistency. Used for session data, caches, and configuration that needs sub-millisecond reads everywhere.

• CONVO_CACHE: Kelly chat sessions with 24-hour TTL — each conversation preserved for context continuity
• VANTAGE_KV: Strategic intelligence cache for the Vantage agent’s signal processing
• CONFIG: Word Orb runtime configuration (feature flags, rate limits, API version routing)
• RATE_LIMIT: Per-key rate limiting for the Word Orb API (sliding window counters)
• WORD_CACHE: Hot word lookup cache to avoid D1 round-trips for frequent queries
• FEEDBACK: User feedback collection buffer for the Word Orb experience

Sovereign AI: Local RTX 5090 Pipeline

Production content is generated on owned hardware — an NVIDIA RTX 5090 (32GB VRAM, Blackwell sm_120 architecture) running the entire voice-to-video pipeline. This was a 3-month breakthrough: PyTorch 2.10.0+cu128 nightly was the first to support Blackwell, unblocking every local AI tool that had failed on this GPU since January 2026.

Voice Pipeline: Chatterbox TTS

• Model: Chatterbox v0.1.7, zero-shot voice cloning
• Speed: 2.3x realtime on RTX 5090 (generates 2.3 seconds of audio per second of compute)
• Voices: 3 age-specific references — Little Annie (child), Kelly (adult), Grace Elder (senior)
• Output: 25 WAV files per lesson (5 tiers × 5 phases), 5.2 minutes total per lesson
• Conda environment: chatterbox

Lip Sync Pipeline: SadTalker

• Mode: crop-512 optimization (eliminates Poisson blending, which was 27s/clip at full resolution)
• Speed: ~45 seconds per clip (1.7x faster than baseline)
• Batch estimate: 365-day lesson set in ~21 hours (single overnight run)
• Compositing: ffmpeg 3-layer assembly (background + static body + animated face + text overlays) in <1 second per clip
• Conda environment: sadtalker

Cinematic Video: Multi-Engine

• Kling 3.0 Pro: $0.112/sec via fal.ai. Face consistency + element control. 1920x1080, H264, 24fps. Tested and proven (classroom, library scenes). ~143 seconds render per 5-second clip.
• LTX-2.3 + ID-LoRA: Free local generation. FP8 weights (29.1GB) + Gemma-3-12b text encoder (24GB) + ID-LoRA weights (1.1GB). Zero-shot from Kelly’s canonical still. Fits in 32GB VRAM at 512x512.
• HeyGen v2 (Genny agent): Talking-head batch pipeline. 1,080 completed, 334 queued. R2 path: kelly/video/en/adult-{archetype}/day-{NNN}/{phase}.mp4
• Sovereign pipeline: 187 videos produced (37 complete days of adult-explorer lessons). Chatterbox + SadTalker + ffmpeg. Performance: 4.0 minutes for 63.7 seconds of finished video.

Kelly’s AI Brain

Kelly’s conversational interface is a separate Cloudflare Worker (kelly-chat) that orchestrates Claude Haiku responses, text-to-speech, lip sync animation, and emotional expression — all via server-sent events (SSE) for real-time streaming.

• POST /v1/chat: SSE streaming conversation — token-by-token response with expression inference in the done event
• POST /v1/chat/complete: Non-streaming completion for batch and programmatic use
• GET /v1/speak: Real-time TTS for any text input
• GET /v1/context: Current lesson context and state for a session
• GET /v1/session/:id: Session retrieval (24h TTL in KV)
• Expression inference: Each response includes an emotion classification (curious, excited, thoughtful, empathetic, amused, engaged) that drives Kelly’s facial animation
• Audio pre-gen: Nightly crons at 02:00 and 03:00 UTC generate and verify 150 audio files per day across 10 priority languages and 3 age groups

Email Infrastructure

The email engine is a Cloudflare Worker with its own D1 database, connected to Resend for delivery. All 6 company domains are verified for sending.

• 6 sending domains: lotdpbc.com (Nicolette), wordorb.ai (Kelly), thedailylesson.com (Kelly), nicoletterankin.com (Nicolette), ilearn.how (Hello), curiouskelly.com (reserved)
• 7 HTML templates: welcome_developer, enterprise_intro, enterprise_followup_1-5, partnership_intro
• 3 sequences: Enterprise cold outreach (5-step), developer welcome (3-step), learner welcome (3-step)
• Brand system: Cormorant Garamond headings, DM Sans body, #1A1A1A ink, #FAF8F5 paper, #B8860B gold accents
• Security: SVIX HMAC-SHA256 webhook verification with 5-minute replay protection
• Batch operations: POST /api/email/sequence/batch-start (100), POST /api/email/contacts/batch (500)
• Dynamic enrichment: Follow-up emails fetch today’s word, IPA, and translations from Word Orb API at send time

Monitoring & Observability

• lotd-healthcheck: 15 endpoints monitored every 5 minutes (6 critical, 5 product, 4 infrastructure). Email alerts via Resend on consecutive failures. Dashboard at lotd-healthcheck.nicoletterankin.workers.dev
• Observability D1: Page view events, beacon data, UTM tracking, operator alerts for high-value events (school lane starts, district lane starts, institutional checklist opens)
• Orwell dashboard: Real-time monitoring of all 15 epic Workers at orwell-dashboard.nicoletterankin.workers.dev — auto-refresh every 60 seconds
• Vantage: Weekly strategic reports with signal scanning daily at noon UTC. Anti-pattern detection catches issues like “distribution without demand” and “breadth without depth.”

Domains & DNS

10 domains, all managed in Cloudflare DNS with orange-cloud proxying where applicable.

• wordorb.ai: Primary product domain. Subdomains: api, docs, mcp, voices
• thedailylesson.com: Consumer flagship. Vercel (Next.js) with Cloudflare Worker overlay. Both bare and www routes active.
• curiouskelly.com: Kelly’s brand/marketing site with Stripe checkout integration
• lotdpbc.com: Corporate site (this site). Cloudflare Worker SSR.
• nicoletterankin.com: Founder’s personal site. Hosts observatory.nicoletterankin.com for Orwell dashboard.
• ilearn.how: Education brand site (deferred). Email verified in Resend.
• wordorb.io, nicoletterankin.org, mynextlesson.com, dailylesson.org: Brand protection and alternate entry points.

The Code Pattern

Every Worker follows the same conventions, making the fleet predictable and maintainable by a single operator.

• Routing: Regex-based path matching in the fetch handler — no framework, no router library
• Responses: jsonResponse(data, status) with CORS headers, safeJSON(str) for safe parsing
• Pagination: parsePagination(url) extracting limit/offset from query params
• CORS: Access-Control-Allow-Origin: * on all API endpoints
• Versioning: {name}_versions table with semver in every database
• D1 binding: Always named DB
• Seed data: INSERT OR IGNORE for idempotent data loading (D1 file import has known bugs with large statements and escaped quotes)

Why This Architecture

A single founder built and operates 80+ production services, 49+ databases, and a complete AI video pipeline. The architecture makes this possible by eliminating operational overhead: no servers to patch, no containers to orchestrate, no scaling to configure. The edge does the work.

It also means no single vendor can shut down the platform. Voice and video production run on owned hardware. Content is cached at the edge globally. Lessons can be served offline through 8 resilience layers. The only dependency that matters is Cloudflare itself — and the CEO (Matthew Prince) has personally introduced us to his VP of Developer GTM. We’re co-authoring a blog post about building this stack.

The cost of this architecture is complexity in deployment and debugging — 80+ Workers means 80+ deployment targets. The benefit is that each service is independent, cold-starts in under 5ms, and can be updated in seconds without affecting anything else. At $0.50 per million requests and $0.75 per million D1 row reads, the economics are extraordinary for a pre-revenue company.