Skip to content
Raj Bhensdadiya

Neyo: AI Agent Engine

An agent that chains 65 tools in one conversation and streams the answer back while it works.

Role
AI & Full-Stack Engineer
Timeline
2 months
Live site
Visit the live site
React UISSERouterPrompt cacheAgent loopSonnetLight workHaiku65 tools
How it works: the React UI streams from a Claude agent over SSE, a router sends the agent loop to Sonnet and light work to Haiku, prompt caching trims cost, and the agent orchestrates its 65 custom tools.
custom tools
65
lower LLM API costs
~90%
OAuth integrations
10
months, built end to end
2

Overview

Production AI SaaS platform: a Claude-powered agent engine in Rails plus the full React/TypeScript frontend, built end to end in 2 months.

The problem

The product needed an agent that could chain dozens of tools without falling over, stream its responses in real time, and connect to a user's third-party accounts, all without LLM costs eating the margin.

What I built

Built a Claude-powered agent with 65 custom tools, multi-step tool orchestration, and real-time SSE streaming, plus a multi-platform activation pipeline whose 10 OAuth integrations connect a user's third-party accounts on their own.

Results

Cut LLM API costs about 90% with prompt caching, context compaction, and multi-model routing: Sonnet for the agent loop, Haiku for the light work.

Decisions

  • Route each task to the right model

    Sonnet runs the agent loop; Haiku takes the light work.

  • Cache and compact the context

    Prompt caching and context compaction, together with that routing, cut LLM API costs about 90%.

  • Stream while the agent works

    Answers stream back over SSE in real time while the agent chains its tools.

Stack

  • Rails
  • React
  • TypeScript
  • Claude API
  • SSE
  • OAuth
  • PostgreSQL

Built by Raj Bhensdadiya, an AI & full-stack engineer in Montreal. The full work history and the other projects have more, and email is the fastest way to reach me.

More work