Azera

An emotionally intelligent AI chat application with a three-layer cognitive architecture: persistent memory, autonomous reflection, and voice output. Built as a multi-service platform rather than a single prompt wrapper.

Architecture

• Axum backend and Svelte frontend, with local LLM inference and voice cloning integrated end to end.
• Polyglot persistence: relational store (CockroachDB), caching, vector storage for semantic recall, and Meilisearch for fast retrieval.
• CUDA image pipeline: Stable Diffusion acceleration for visual generation alongside text and speech.

What we learned

Azera stress-tested the boring parts: schema evolution across services, cache coherency, vector + keyword hybrid search, and operability when models are slow or OOM. That experience feeds directly into how we design production AI adjacency for clients who need more than a demo.

Repo

Runtime deployments are personal-lab scale, not a hosted product offer.

• GitHub

  Public code and docs for the Azera platform.

• Wordkeep Wiki lab

  Sanitized architecture notes in the demo corpus (projects/azera/).