Kelly.

Kelly teaching Day 1 — “Hook” phase, Explorer archetype

The Five-Phase Lesson

Every day, Kelly delivers a structured lesson built around a single word. The lesson follows five phases, each designed to engage a different mode of learning. This isn’t a flashcard — it’s a complete learning experience that takes 3-5 minutes.

Four Learning Tracks

• Learn: Core vocabulary building — the default track for new learners
• Grow: Advanced vocabulary with emphasis on etymology and cross-language connections
• Teach: Pedagogical track — teaches you how to teach the word to others
• Trivia: Fun-focused track with surprising facts, word games, and cultural connections

Content Scale

• 74,008 base lessons across 19 languages
• 164,250 archetype variants in English — same word, same day, personalized for 10 different learning styles
• 24,989 assessment items in 4 formats across 6 languages
• 365 days of complete content in 8 languages (English, Spanish, French, German, Portuguese, Japanese, Korean, Chinese)

The 10 Teaching Archetypes

Kelly doesn’t teach everyone the same way. She adapts her personality, vocabulary, examples, and tone based on 10 teaching archetypes — each one a different cognitive and emotional approach to learning. Learners select (or are matched to) an archetype, and every lesson is personalized accordingly.

Three Ages of Kelly

Kelly exists across three life stages, each with a distinct voice, visual appearance, and teaching tone. Learners choose the Kelly that resonates with them — or experience all three through the age-transition feature.

Real-Time Conversation

Kelly isn’t just a lesson player — she holds real-time conversations. The kelly-chat Worker orchestrates Claude Haiku responses with text-to-speech and emotional expression inference, all streamed via server-sent events (SSE) for real-time interaction.

How It Works

1. You speak or type — input goes to the kelly-chat Worker
2. Claude Haiku generates a response — streamed token-by-token via SSE
3. Expression inference — each response is classified as curious, excited, thoughtful, empathetic, amused, or engaged
4. Text-to-speech — Kelly speaks her response using her age-appropriate cloned voice
5. Lip sync animation — SadTalker drives facial animation synchronized to the audio
6. Session preserved — conversation context stored in KV with 24-hour TTL for continuity

Conversation API

• POST /v1/chat — SSE streaming conversation with token-by-token response and expression inference
• 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)

Video Production: 2,300+ Videos Across 4 Engines

Kelly’s video lessons are produced across four video generation engines, each with different quality, cost, and identity characteristics. The kelly-video-factory Worker orchestrates all four with a unified production queue.

Media Library

• 15,542 total media files in R2 storage at kelly-assets.nicoletterankin.workers.dev
• 2,300+ lesson videos — HeyGen batch + sovereign pipeline + cinematic tests
• 850 legacy video clips — earlier production experiments
• 36 visual looks across 10 archetypes and 3 age groups
• 5 email avatar states × 3 sizes for email template personalization
• 7 source videos (62MB) providing Kelly’s canonical appearance and voice reference

R2 Video Path Convention

kelly/video/{lang}/{age}-{archetype}/day-{NNN}/{phase}.mp4

Example: kelly/video/en/adult-explorer/day-001/hook.mp4

Sovereign Voice Pipeline

Kelly’s voice is generated on owned hardware — an NVIDIA RTX 5090 (32GB VRAM, Blackwell architecture). No third-party TTS service. No API calls. No per-character costs. The voice references were created from Kelly’s canonical source videos and will never expire, get rate-limited, or be discontinued.

Voice Generation

• 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)
• Output per lesson: 25 WAV files (5 tiers × 5 phases), 5.2 minutes of audio total
• Batch capacity: 365-day lesson set in ~10 hours of voice generation
• Conda environment: chatterbox

Lip Sync Animation

• Model: SadTalker (crop-512 optimization)
• Speed: ~45 seconds per clip (1.7x faster than baseline by eliminating Poisson blending)
• Compositing: ffmpeg 3-layer assembly (background + static body + animated face + text overlays) in <1 second per clip
• Batch capacity: 365-day lesson set in ~21 hours (single overnight run)
• Conda environment: sadtalker

Audio Pre-Generation

Nightly cron jobs (02:00 and 03:00 UTC) pre-generate and verify 150 audio files per day across 10 priority languages and 3 age groups. This means Kelly’s voice is ready and waiting before the learner wakes up — no generation latency at lesson time.

The Player

Kelly’s video player is built into the Daily Lesson consumer site. It auto-plays the video for the current lesson phase, falls back gracefully when video isn’t available, and provides enrichment features during playback.

• Auto-play: Video starts immediately for the current phase (muted for browser compliance, with unmute button)
• Fallback: When no video is available, shows Kelly’s canonical still image with text content
• “Watch Kelly” toggle: When video exists but isn’t auto-playing, a toggle button lets the learner choose
• Visual aids: Word enrichment data displayed alongside the video (translations, knowledge graph, pronunciation)
• Phase navigation: Automatic advance between the 5 lesson phases with video for each

Kelly’s 8-Layer Resilience System

Education infrastructure that requires always-on connectivity excludes the people who need it most. Kelly’s shield system (kelly-shield Worker at shield.kellyai.com) provides 8 layers of progressively degraded but functional lesson delivery — from full-featured edge delivery down to physical books.

1. Layer 1 — Cloudflare Edge: Full lessons with video, audio, chat, and assessments. 330+ locations, sub-50ms latency.
2. Layer 2 — R2 Object Cache: Pre-rendered lesson bundles. No database queries needed. Survives D1 outages.
3. Layer 3 — KV Bundle: Compressed lesson packages in globally-replicated KV store. Fastest reads, smallest payloads.
4. Layer 4 — Service Worker PWA: Full offline capability. Lessons cached on device after first load. Works with zero connectivity.
5. Layer 5 — QR Print: Scannable QR codes on printed materials link to cached lessons. Bridges physical and digital.
6. Layer 6 — SMS Gateway: Text-based lessons for feature phones on 2G networks. Reaches the bottom billion.
7. Layer 7 — Satellite Filecasting: Broadcast delivery via Knapsack for Hope (Galaxy 19 → Americas) and Toosheh (YahSat → Iran/Middle East). No internet required at all — content received via satellite dish.
8. Layer 8 — Physical Book: Print-ready lesson compilations generated from the content engine. Works during total infrastructure collapse.

Where to Find Kelly