Nexus-AI
A full-stack agentic AI platform that generates, runs, and iterates on code inside isolated sandboxes — across persistent sessions.
Case study
Nexus-AI
Problem
What gap does this fill?
Most AI coding tools stop at generating a code snippet. There was no lightweight platform that could generate a complete project, run it in a sandboxed environment, and let a developer iterate on it across sessions — with full session persistence and a durable workflow that survives failures.
Solution
What was built
Nexus-AI is a full-stack platform built on Next.js, tRPC, Prisma, and PostgreSQL, powered by a 3-stage durable agent workflow using Inngest and Google Gemini. The agent can generate files, run terminal commands in an E2B sandbox, and iterate autonomously for up to 15 tool-call cycles per session.
My Contribution
What I specifically built
I designed and built the complete platform — from the type-safe tRPC API layer and Prisma schema through to the Inngest-powered durable workflow and Gemini tool-calling integration. The E2B sandbox setup and the session persistence system across the PostgreSQL-backed project history were the two most technically involved pieces.
Technology
Stack used
Frontend
Backend
AI / Agents
Infrastructure
Engineering Decisions
Notable technical choices
Durable 3-stage agent workflow via Inngest
ensures the workflow survives server restarts and network failures mid-generation.
Up to 15 autonomous tool-call iterations per session
allows the agent to self-correct and iterate rather than stopping at a single generation pass.
Sandboxed code execution via E2B
runs generated code in a completely isolated environment, preventing any host-system exposure.
Type-safe end-to-end API with tRPC
eliminates an entire class of runtime errors at the frontend/backend boundary.