Back to work

delivXchange | AI native platform

I am in build mode on delivXchange. It is AI native from the start. We use agents, voice, MCP, RAG, and agentic deployment. This is the current technical shape.

Agents as the core

We model work as tasks an agent can own. Each agent has a clear goal like intake, routing, operations, or support. Agents can call tools and talk to each other when needed.

Voice agent

Real time speech in and speech out with streaming. The voice agent hands off to other agents when the call needs action. We keep latency low with lightweight turns and partial results.

Model Context Protocol

We use MCP to plug tools and data into agents in a standard way. Tools include internal APIs, data fetchers, vector stores, and job runners. MCP lets us add or remove tools without rewiring the whole system.

Retrieval augmented generation

We use RAG to ground answers in our data. Content is split, embedded, and stored in a vector store. We track versions. We add safety rules so the agent uses only allowed sources.

Agentic flows and planning

We let the system plan a few steps ahead when useful. Plans are short and checked at each step. Agents can reflect and try a different path if the result is weak.

Deployments and evaluation

We ship agents behind feature flags. We run offline evals on tasks and also shadow live traffic. We log inputs and outputs with scores. We keep a feedback loop for humans in the loop.

Operations and safety

We monitor token use, latency, and error rates. We set guardrails for PII, tool scope, and rate limits. We prefer simple fallbacks over complex retries.

Stack idea

TypeScript and Node for services. React for ops UI. WebSockets or Server Sent Events for live updates. Postgres for source of truth. Vector store for embeddings. Queue for jobs. Infra with containers and CI and CD. This can change as we learn.

Value

The goal is simple. Fast intake. Clear routing. Less back and forth. A system that learns from real work.