Voice and chat agents
Live conversational pipelines: audio over LiveKit, Gemini Multimodal Live (barge-in, low latency), and tool calling against real accounts. On Rembrr that includes Gmail, Google Calendar, Drive, cited web search, and tasks.
I am Andrés Navarro Gómez, a Backend & AI Engineer. I design and build APIs and conversational systems that ship — live voice, tools, memory — not one-off demos. The public case is Rembrr: the voice and agent backend of a multimodal platform on the App Store.
Live conversational pipelines: audio over LiveKit, Gemini Multimodal Live (barge-in, low latency), and tool calling against real accounts. On Rembrr that includes Gmail, Google Calendar, Drive, cited web search, and tasks.
The assistant remembers across conversations: embeddings over pgvector, hybrid ranking (dense + lexical), and fact extraction off the request path so it does not add latency to voice.
APIs and services in Python (FastAPI) and Node.js/TypeScript (NestJS): multi-tenant systems, queues, integrations, and the wiring an agent needs to take real actions.
This is not a fixed-price package. It is the order I have used to take voice and agents to a shipped product (Rembrr) and backends to platforms in use.
Scope voice vs chat, tools, latency, and sensitive data. Without that the model stays a demo you cannot operate.
A thin live path: one conversation, a few tools, latency and task-completion measured. On Rembrr the core was the pipeline that holds a live conversation and dispatches tool calls.
Memory/RAG, token encryption (on Rembrr: KEK/DEK envelope with Cloud KMS), explicit connection lifecycle, CI/CD. Shipped to the App Store; ongoing work is latency and session stability.
Need a production backend or a voice/chat agent that actually does the work? Email me. No form.
Email andresnavarrodev@gmail.com