Products
Three surfaces.
Production tooling, foundational mobile-AI patterns, and a public sandbox.
cuppa.studio
Our production AI tool. An online presentation IDE where a prompt becomes a complete animated, narratable deck. Live editor and real-time preview side-by-side; code blocks with syntax highlighting, mermaid diagrams, charts, terminal simulations; AI-generated narration via Auto Drive. Share with a link — anyone can view, no account needed.
Decks also export as a self-contained .cup
file with a <cuppa-cue> web component
for embedding elsewhere.
Cuppa Harness
The library of design patterns and primitives we use to build AI on mobile — cross-platform across iOS, Android, and React Native. The fundamentals that don't break when the model below them changes. Now powering the Cuppa app end-to-end.
What's in it
- Provider-agnostic protocols — swap or fan out across Claude, Gemini, Groq, Ollama Cloud, and Apple Foundation Models without rewriting call sites.
- Multi-provider orchestration — Council fan-out, moderator-as-judge, smart routing by query class and live cost. Cheaper, faster, surfaces a "verify" badge when models disagree.
- Device-context awareness — battery, thermal, network quality, and foreground/background lifecycle. The piece cloud SDKs structurally can't see; the harness adjusts routing and streaming based on what the phone is actually doing.
- First-class imaging lane — image input, on-device CV preprocessing (OCR, faces, scene, PII heuristics), default-on privacy redaction with a visible toggle, vision-aware routing, and Council-for-vision with visual-hallucination signals.
- Safety primitives — on-device PII redaction before send, prompt-injection detection with a defensive wrap, and pluggable output filters that catch leaked secrets before render. One consistent shield-icon surface across input and output.
- Memory & context patterns — conversation buffer, summarisation rolloff, retrieval scaffolding.
- Output repair — fence stripping, brace block extraction, JSON candidate ladders, reasoning-tag handling. Took us from ~60% to 100% schema-valid output on on-device models.
- Cost & eval observability — per-reply token usage and dollar cost on every model card, failure-mode taxonomy, and a validity dashboard for tuning.
Designed for non-breaking growth: tool calls, additional multimodal kinds, and per-vertical privacy policies all land as 1.x extensions without changing the 1.0 protocol shape.
mycuppa.io
Our open-source R&D sandbox. Experiments, prototypes, and research notebooks tied to GitHub. Released as we explore, before they're polished — lower barrier, faster ship velocity. The featured ones graduate to a spotlight here.