OurStoryTime is a collaborative writing platform where children write stories together, generate stories from photographs of their own drawings, hear them narrated aloud, and read each other's finished work in an animated 3D flip-book. The 3D reader exists as motivation: making other kids' work feel worth reading is what makes writing feel worth doing.

The AI layer runs on LangGraph rather than a single-shot chain. A StateGraph loops between reasoning and tool invocation through a ToolNode until the model is done, and MemorySaver persists conversation state across turns so a session survives page reloads. Anthropic's Claude drives reasoning and tool selection, with the prompt-caching beta header enabled against a long static system message, and trimMessages bounding context-window growth so cost does not scale with session length. Tools live in an externally hosted wxflows registry, which means new agent capabilities can be added without redeploying the application.

Provider routing is a deliberate architectural decision rather than an integration detail. A unified gateway routes dynamically between Anthropic and Google instead of exposing one provider's capabilities as tools to the other. That choice was made after researching production multi-LLM patterns; it avoids vendor lock-in, enables cost-based model routing, and prevents the tool-coupling debt that shows up eighteen months later. Token-by-token SSE streaming runs on the Next.js edge runtime for low first-token latency, and every provider key is proxied server-side so no credential reaches the browser.

AES Connect is a mobile platform for the Alliance of Sahel States (Burkina Faso, Mali, Niger) and its diaspora. Governments post opportunities and reach citizens directly, young people find work and projects, and diaspora members connect back to the communities they came from. It is over 200 components across three nested tab environments, with context-aware navigation modeled on super-app patterns so users move between civic, marketplace, and services contexts without leaving the application.

Multi-nation data isolation is enforced at the database layer using PostgreSQL Row Level Security, so community-scoped content cannot leak across member states regardless of what the client does. This is a correctness requirement, not a hardening pass: a client-side filtering bug in this application would be a cross-border data exposure.

State persists offline through Redux Persist and AsyncStorage because connectivity in the target markets is intermittent rather than reliable. The application has to be useful before the network returns, which makes offline-first the default architecture rather than an optimization applied later. Production release to the App Store and Play Store is in preparation via EAS Build.

Faso Kuma is a voice-first translation application for French, Mooré, and Dioula. Voice-first is an architectural conclusion, not a preference. The market has roughly 85% mobile penetration against 19.9% internet penetration, and a text-first interface would exclude exactly the people the product exists to serve.

The model layer is built on open-source work that can be fine-tuned rather than rented: Whisper for speech-to-text, Meta's NLLB as the translation base for low-resource languages, Wav2Vec2 for speech representation, and MALIBA-AI's Bambara-specific models. A caching layer sits in front of third-party translation APIs so that interim external results populate a first-party dataset, progressively reducing dependency as coverage grows.

The strategic core is the dataset. Crowdsourced translations are promoted to confirmed status only after three independent validations, stored in segregated tables, and licensed CC BY-NC-SA, open to African researchers and universities, with commercial rights reserved. Building a corpus for languages that have almost no digital tooling, and keeping it under African control, is the point of the project.