Personal Project|2025

OrbitAgent

Multi-Tool AI Agent Platform

Full-stack AI agent that connects to GitHub, Linear, and Slack via OAuth and performs real work across them with human-in-the-loop approvals and a full audit trail. Built with Next.js 15, FastAPI, SSE streaming, and multi-provider LLM orchestration (Claude + Gemini).

Role:Full Stack Engineer

OrbitAgent — Multi-Tool AI Agent Platform built by Adnan Ali

The Challenge

Knowledge workers juggle 10+ SaaS tools daily — GitHub for code, Linear for tickets, Slack for comms — and context-switching is expensive. Generic LLM chatbots hallucinate because they can't see real data, and giving an AI unsupervised write access to production tools is unsafe. A trustworthy agent needed scoped OAuth, human approvals before any mutation, and a complete audit trail.

The Solution

Built a monorepo platform with a Next.js 15 App Router frontend and an async FastAPI backend. The LLM streams responses over Server-Sent Events and can call registered tools across connected services in one thread. Any write action (send message, create issue, close PR) pauses execution, emits an approval_required event, and waits for a JWT-signed approve/edit/reject decision before resuming. OAuth tokens are Fernet-encrypted at rest, every tool call is persisted as a first-class DB row, and workflows support Liquid-style {{variable}} templates for reusable agent runs.

Key Features

Streaming multi-tool AI chat over SSE with Claude + Gemini orchestration

Human-in-the-loop approvals for every write action (Approve / Edit / Reject)

OAuth connections for GitHub, Linear, and Slack with Fernet-encrypted token vault

Workflow engine with {{variable}} substitution and one-click reruns

Full audit log — every tool call recorded with params, result, and deep-link back to conversation

Multi-tenant workspaces with memberships, roles, and workspace-scoped queries

Results & Impact

Shipped a production-grade agent platform on Railway with ~27 passing tests and 42% backend coverage across 22 Next.js routes. The human-in-the-loop safety layer makes the agent safe to wire into real production tools, and the audit-first design means nothing the agent does is invisible.