Assistant Core is the cloud runtime layer for voice, reasoning, memory, knowledge base, and tool calling across web, apps, smart devices, IoT, and robots.
Once your assistant needs to listen, speak, remember, call tools, understand private data, and run on real devices, teams end up stitching many infrastructure layers together.
A natural voice experience needs VAD, ASR, LLM, TTS, audio streaming, interruptions, and latency control in one pipeline.
The assistant must call APIs, read data, trigger workflows, control devices, and extend through MCP or governed tool calling.
Smart devices, robots, and IoT hardware should keep microphone, speaker, sensors, actuators, and connectivity while the cloud runtime handles reasoning, memory, and voice.
Multi-tenancy, auth, permissions, conversation history, memory, and observability need to ship with the assistant from day one.
Assistant Core brings the hard parts of voice, agents, tools, and device integration into one assistant-level architecture.
Explore runtimeShape the prompt, model stack, knowledge base, long-term memory, brand voice, domain, and behavior for each assistant.
Customize the voice pipeline from VAD, ASR, LLM, and TTS to realtime speech-to-speech for web, apps, or screenless devices.
Connect the assistant to APIs, data, workflows, internal systems, smart homes, or robots through tool calling and MCP.
Embed the assistant into products through Chat API, Voice WebSocket, MQTT gateway, widgets, or custom hardware.
Manage assistants, users, conversations, memory, devices, access control, and production operations.
Assistant Core fits products where assistants need to speak, remember context, call tools, and connect to software or devices.
Integrate an assistant into landing pages, apps, docs, dashboards, or SaaS products through API, WebSocket, and embedded widgets.
Build internal assistants for support, sales, HR, training, knowledge management, and workflow automation on private data.
Turn expertise into a virtual assistant for students, communities, customers, or followers with private memory and knowledge base.
Bring voice assistants to education robots, smart toys, smart home devices, learning hardware, kiosks, or compact AI hardware.
Assistant Core is more than a chat UI. It provides the layers needed for private knowledge, memory, tool calling, realtime voice, and safe production operations.
Control system prompt, model stack, knowledge base, tools, MCP connections, behavior, and access for each assistant.
Connect documents, websites, and structured sources so answers stay tied to the information your team trusts.
Voice pipeline VAD → ASR → LLM → TTS with audio streaming for natural spoken conversations.
Let the same assistant call built-in tools, custom APIs, server MCP, endpoint MCP, and device-side MCP tools.
Customize interface, fonts, emoji, wake word, and backgrounds, then deploy to smart devices, IoT hardware, or robots.
The assistant remembers context across sessions, extracts important facts, and retrieves them through pgvector when needed.
Mix models for reasoning, ASR, TTS, realtime voice, or embeddings based on each assistant's needs.
Provider availability depends on deployment configuration and workspace API keys.
Voice, memory, tools, and device integrations all pass through authentication, authorization, tenant isolation, and operational observability.
TLS 1.3 encryption on all HTTP, WebSocket, and MQTT connections. No plaintext data in transit.
JWT access + refresh tokens, Google/GitHub OAuth with HMAC-signed state and Redis-backed nonce to prevent replay attacks.
Role-based access control with scoped JWT claims. Per-assistant permissions and multi-tenancy isolation at the request level.
Encryption at rest via PostgreSQL and S3 server-side encryption. Row-level isolation between tenants.
Per-user and per-IP rate limiting. CORS policies. DDoS mitigation at the load balancer layer.
Request logging, conversation-level LLM/tool traces in the admin dashboard, and assistant metrics for security monitoring.
Technical deep-dives, product updates, and best practices for building AI-powered products.
Explaining Acoustic Echo Cancellation from the NLMS algorithm to real-world deployment in the browser (AEC3) and on ESP32 (ESP-ADF). Includes a comparison of AEC support across OpenAI, Gemini Live, and xAI Grok.
How we built a retrieval-augmented generation pipeline handling PDF, DOCX, and web URLs with chunking, embedding, and cosine search — reaching 91% accuracy compared to 34% with a plain LLM.
Why we dropped WebSocket for hardware devices and moved to MQTT + UDP — cutting latency from 380ms to 165ms, RAM from 45KB to 12KB, and gaining 40% battery life.
How we integrated OpenAI, Anthropic, Google, xAI, DeepSeek, and Xiaomi through native SDKs — letting users switch models mid-conversation, saving up to 90% on input-token costs with prompt caching, and deploying with zero downtime.
Start with one assistant, then configure prompt, voice pipeline, memory, knowledge base, tool/MCP connections, and deployment surfaces for your product.