Roadmap & status
Working software first, then deeper
The strategy: get a genuine Socratic conversation working end-to-end fast, then improve every layer — inference, speech, hardware, pedagogy — independently. Each phase below produces something a child can actually use.
The phases
Phase 0 — Proof of pedagogy Complete
A text-mode Primer that holds a genuine Socratic conversation on any computer. The exit test — a 15-minute conversation that asks more than it answers, catches parroting, suggests breaks, and remembers last time — is met.
- Streaming conversation with session persistence, resume, and long-term memory
- Curated, licensed children's knowledge corpus (English + German) with tuned hybrid retrieval
- Engagement, concept, and comprehension classifiers feeding a persistent learner model
- Spaced-repetition vocabulary and session-break suggestions
- Desktop app, multilingual prompt packs (English/German production, Hindi preview)
Phase 1 — Local inference In progress
Run the conversation loop offline on hardware families own. Target: under three seconds to first token on at least one local platform.
- Embedded local inference on laptop/desktop — landed; device benchmarking under way
- Qualcomm NPU backend — the Primer now holds a full multi-turn conversation on a Snapdragon 8 Elite phone's Hexagon neural processor at ~25.7 tokens/s, first token in under a second (June 2026)
- Per-turn hybrid routing between a local model and an optional cloud model — landed, opt-in
Phase 2 — Speech In progress
Talk to the Primer instead of typing. Much of this landed ahead of schedule.
- Offline voice loop (voice detection → transcription → response → synthesis) — working today
- Strict turn-taking: no barge-in either direction, by pedagogical design
- Native Apple speech engines on macOS; multilingual synthesis incl. Hindi in evaluation
- Ahead: ambient-noise robustness, echo cancellation, voice-profile selection
Phase 3 — A device a child can hold Started
The Primer as a physical object: turn it on and talk, no other equipment. Android app packaging — the deployment path to phone-class hardware — is working, and the app now runs the conversation on the phone's neural processor.
- Android app builds and installs as an APK carrying the NPU runtime; the Primer holds a coherent multi-turn conversation on-device, generating every token on the phone's neural processor
- Ahead: display (colour e-ink or repurposed tablet), microphone array and speaker, battery and drop-resistant enclosure
Phase 4 — Pedagogical depth Ahead
From a Socratic chatbot to a genuinely effective long-term learning companion.
- Curriculum alignment (Australian Curriculum, IB PYP)
- Multi-session learning arcs that build over weeks
- Read-only parental insight — never surveillance
- Collaborative mode: two children sharing one Primer
- Opt-in, parent-consented, anonymised language corpus to improve child-calibrated models — with on-device scrubbing before anything leaves
Dated milestones
Full text-mode Primer validated end-to-end on an Android phone (Snapdragon 8 Elite) — conversation, persistence, and the complete classifier chain.
The neural-processor inference pipeline validated on a Snapdragon 8 Elite phone — a 4-billion-parameter model generating coherent text on the phone's Hexagon NPU, comfortably inside thermal limits.
Android app packaging landed: the Primer's desktop app builds and runs as an Android APK, carrying the NPU runtime — the path to the first fully self-contained phone deployment.
First token generated on the phone's neural processor by the Primer's own app — the NPU backend running end-to-end inside the Android build, not a reference proxy.
A full multi-turn Socratic conversation runs on the phone's neural processor — near-instant, stable across turns, with a responsive phone-screen layout. Entirely on-device.
On-device throughput measured inside the app: ~25.7 tokens/s sustained, first token in under a second — well past the target for comfortable conversation.