An AI assistant sold as a product,
with a neural router that learns from its own telemetry
Our own product sold to third parties: a FastAPI backend with a single-page app, a memory engine with a 4096-dimension sparse representation and two-level memory, a model cascade, image generation, a Telegram bot, API-key auth with rate limiting, and a Kolmogorov-Arnold network router written in pure numpy that chooses the memory mode and retrains on telemetry with a fallback to thresholds. 18 of 18 intelligence tests passed; router accuracy 0.91.
What it is
A packaged AI assistant with documentation that a buyer can run: a brain engine with a sparse distributed memory, a cascade of models from small to large, image generation, a Telegram bot with keep-alive, and API-key authentication with rate limits.
The unusual part is the router: a real Kolmogorov-Arnold network in pure numpy with analytical backpropagation that picks the memory mode instead of hard thresholds, retrains on its own telemetry, and falls back to thresholds if it degrades. A kill switch and an admin key guard unauthorised operations.