Zahara AIhttps://zahara-v1.vercel.app

The Problem

Zahara is a multi-tenant platform for building AI agents through several distinct surfaces — a conversational "Vibe" builder, a visual "Flow" builder, a full-IDE "Pro" builder, and a guided "Clinic" builder. As usage grew, a set of UX and trust issues surfaced across these builders, and the Pro IDE carried a subtle but serious data-scoping bug: its file API loaded files across agents instead of scoping them to the active one, so a user could see the wrong agent's files. In a multi-tenant AI product, that kind of cross-tenant bleed erodes trust quickly.

The Approach

I traced the Pro IDE issue to its root cause — an unscoped file-loading endpoint — and refactored file handling end to end so every file operation is scoped to the active agent. That spanned both the FastAPI backend and the Next.js front end (the file tree, the client store, the page, and the typed API layer), keeping the contract consistent across the stack. In parallel I worked through seven client-reported UX and trust issues across all four builders, and ran the changes through three review rounds before handing back clean patches against the client's branch.

Architecture

Zahara pairs a Next.js App Router front end with a FastAPI (Python) backend. The front end exposes four builder surfaces over a shared component and state layer; the Pro builder is a full in-browser IDE with a file tree backed by a client-side store and a typed API client. The backend serves a per-agent file API and the multi-tenant agent runtime. The refactor established a clear invariant: file reads and writes are always scoped to the active agent, enforced on the server and reflected consistently through the client store and UI.

Outcome

The Pro IDE file-scoping refactor closed the cross-agent file bleed across five files on both the backend and front end, with the agent-scoped contract verified over three audit rounds. The seven client-reported issues spanning the Vibe, Flow, Pro, and Clinic builders were resolved, and the work shipped as clean, reviewed patches applied to the client's branch — measurably improving the reliability and trustworthiness of the builder experience.

Tech Stack