Edward
Edward is prompt-to-production execution infrastructure: admission-controlled BullMQ runs, resumable execution over an append-only event log, per-chat Docker sandboxes, and S3 + Cloudflare previews.
Overview
Most AI builders stop at generation. Edward is everything after. A user describes an app in plain language, and the system takes it from prompt to running software: generate code, inspect and edit files, execute it in an isolated environment, publish a live preview, and sync the result back to GitHub without leaving the product.
Most AI builders stop at generation. Edward is everything after. A user describes an app in plain language, and the system takes it from prompt to running software: generate code, inspect and edit files, execute it in an isolated environment, publish a live preview, and sync the result back to GitHub without leaving the product.
The system is made up of a Next.js web app for chat and UI, an Express API for managing runs and previews, shared packages for auth and data access, and Docker-based environments for running generated apps.
Behind the scenes, it uses Postgres for product data, Redis queues for long-running work, Docker for isolated runs, and AWS plus Cloudflare for shareable previews. That makes the output something you can actually use and build on, not just model text in a chat window.
How It Was Built
The main technical choices behind the product, from system design to the parts that make it work day to day.
- Designed run admission control: Postgres advisory locks enforce global, per-user, and per-chat limits before a BullMQ worker begins work.
- Made generation resumable with an append-only event log: typed stream events persist, and reconnecting clients replay missed file, tool, build, and preview state instead of restarting a run.
- Designed run admission control: Postgres advisory locks enforce global, per-user, and per-chat limits before a BullMQ worker begins work.
- Made generation resumable with an append-only event log: typed stream events persist, and reconnecting clients replay missed file, tool, build, and preview state instead of restarting a run.
- Isolated generated-code execution in per-chat Docker sandboxes with Redis-coordinated provisioning, resource and network limits, and an export path to S3-backed previews behind Cloudflare routing.
- Added a validation gate before generated code ships: syntax (node --check), type (tsc --noEmit), and production build checks run inside the sandbox, and failures feed a retry prompt back to the model.
Impact
- Turned prompt-to-application into a single workflow: generate, edit, run, preview, and sync back without switching tools.
- Removed blocking build and preview steps from the chat loop by handling them asynchronously, so the conversation stays responsive.
- Turned prompt-to-application into a single workflow: generate, edit, run, preview, and sync back without switching tools.
- Removed blocking build and preview steps from the chat loop by handling them asynchronously, so the conversation stays responsive.
- Kept the output as a real repo with shareable preview links, so generated work can be reviewed and continued in normal engineering workflows.
Highlights
- Built around execution, not just generation.
- Combines product design with the backend systems needed to actually run apps.
- Built around execution, not just generation.
- Combines product design with the backend systems needed to actually run apps.
- Hands off cleanly into normal code review and engineering workflows.
Tech Stack
More Projects
Additional work across AI products, developer tooling, and full-stack systems.
Agentic Chat
An AI chat platform that routes each request through the right context — memory, documents, tools, or research — and acts on the answer through connected apps.
Bonkers by Foyer
Creative production system built at Foyer Tech — led the v2→v3 rebuild with reusable templates, multi-model routing, and faster repeat workflows for high-quality visual asset creation.
Fieldcraft
A careful Chrome extension copilot that reads the open job tab, evaluates fit against your profile, researches the company, and drafts truthful answers for review before any form is filled. It never auto-submits.