Loading...
Back to Portfolio
In Progress2026
Bert
Local-First AI Companion
100% local Super-AI companion for solo founders. Sub-second voice loop, 26 callable tools, and federated memory across Discord, mobile, and dashboard.
AI/MLPythonVoiceLocal LLM
Overview
Bert is a privacy-first AI companion that runs entirely on-device. An MLX-powered local LLM, Whisper STT, and Piper TTS deliver sub-second voice response without sending a single token to the cloud during normal operation.
Beyond chat, Bert exposes 26 callable tools — calendar, email, code search, memory queries, web ask — gated by an 8-class authority matrix and approval queue. A LightRAG knowledge graph federates episodic and semantic memory, with a daily consolidator that extracts personality traits and surfaces relationships across people, calendar, and email.
Three coordinated surfaces share a single FastAPI brain: a Discord-native voice + text channel as the primary transport, a React Native mobile cockpit (Expo, Zustand, MMKV with offline outbox + DLQ), and a React + Vite operator dashboard for cost guardian, ingest control, and live diagnostics. Currently in pre-launch — targeting late May 2026.
Role
Solo Architect & Engineer
Technologies
27+ tools
Year
2026
Key Features
Sub-Second Voice Loop
314–750 ms time-to-first-audio (warm) via tools-trim and context-shrink. Whisper in, Piper out, MLX brain in between.
26 Callable Tools
Read, capture, calendar, email, web ask, web browse, code review, agent invocation. Authority matrix gates writes via approval queue.
100% Local LLM
MLX Qwen 14B primary, Gemini fallback only for web ask. No tokens leave the device for normal operation.
Federated Memory
Episodic + semantic split with LightRAG knowledge graph, daily consolidator, and per-person CRM cards across calendar and email.
Multi-Surface Cockpit
Discord-native voice + text, React Native mobile app (offline outbox, durable persistence), and React/Vite operator dashboard.
Pre-Meeting Brief
T-5 min calendar hook generates a person + topic briefing with anti-fabrication guards. 38 unit tests on the prompt path.
Proactive Daemon
Rule loop with quiet hours, frequency caps, persistent dedupe, and a 09:30 journal nudge. Bert reaches out, not just reactive.
Authority Matrix
8-class permission system (a–h) with approval queue, audit log, rate limiting, and trait decay. Writes never slip past the gate.
Tech Stack
Brain & Runtime
MLX (Apple Silicon)Qwen 14BWhisper STTPiper TTS
Backend & API
PythonFastAPIWebSocketServer-Sent EventsSQLite
Memory & Retrieval
LightRAGEpisodic + Semantic splitTrait extractionRelationship graph
Mobile (React Native)
ExpoZustandMMKVReanimated 3Jest + RNTL
Dashboard
React 19ViteTypeScriptFramer Motion@dnd-kit
Integrations
Discord (sole transport)Outlook MCPGemini (web fallback)Playwright
Interested in a similar solution?
Let's discuss how I can build something like this for your business.