The original idea
Nina has been thinking about a prompt-based software building and customization factory since roughly 2021: software made for one person, team, or moment without first turning the idea into a conventional software project.
WHY + HOW
WELCOME TO THE MONGODB BUILD LAB
(A CUSTOM SOFTWARE FACTORY)
One prompt becomes a deployed app with a MongoDB Atlas backend and a built-in LLM, without asking the participant to assemble the repository, infrastructure, tests, or release process.
WHY
Remove the work between a prompt and a deployed app.
A coding agent is not a release process
A prompt still needs intake, planning, constraints, database and model access, tests, repair, deployment, tracking, and operations before it becomes a reliable live app.
Put the environment around the agent
The Build Lab gives the coding agent a constrained path from an attendee's idea to a live application, then checks the result before anyone receives the link.
TIME + SCALE
What was built in the 15-day window.
15 days
The repository began on July 29, 2026. The event is August 13, 2026.
Nina + Justin
Coding agents helped write and revise the software. Nina and Justin made the product, system, constraint, evaluation, testing, and event-operation decisions.
1,307 nonblank lines
Measured August 11 across the artifact checks, runtime/browser verification, and production end-to-end runner. This excludes tests, provider adapters, worker orchestration, UI, and APIs.
STACK
The pieces behind the factory.
Static web + React
Static HTML, CSS, and JavaScript run the event surfaces. Generated apps use React and Vite.
Node.js services
Node.js runs the APIs, worker, email jobs, generated-app runtime, verification, and publication flow.
MongoDB Atlas + Voyage AI
Atlas stores platform and app data and provides search, vector search, feedback, voting, telemetry, and judging. Voyage AI provides embeddings.
Model gateway, agents, Vercel
Model calls go through MongoDB's internal gateway. Coding-agent adapters build the apps. Vercel and Vercel Blob serve the platform and published artifacts.
LOOSE ARCHITECTURE
One durable record connects every stage.
1
Intake and safety checks create the MongoDB Atlas submission record.
2
A planning agent produces a focused brief for attendee review or automatic approval.
3
The worker gives a coding agent an isolated starter app, narrow runtime API, and explicit rules.
4
Source, manifest, browser, and Atlas-backed behavior checks drive bounded repair or fallback.
5
The platform creates Atlas data and indexes, publishes the artifact, and smoke-checks the hosted app.
6
Atlas updates feed tracking, email, Apps, projection, organizer, judging, and feedback surfaces.
Generated browser apps never receive database credentials or model keys. They call a narrow runtime API; the trusted platform validates and performs each Atlas or model operation.
HARNESS STEPS
Build, check, repair, publish, check again.
1
Normalize the prompt, identity, sharing choice, and event metadata.
2
Run deterministic and model-assisted safety checks.
3
Create and approve a focused build brief.
4
Build from the constrained starter and declared Atlas data model.
5
Check source, manifest, requested capabilities, and visible user journey.
6
Run browser tests, including a real Atlas-backed mutation.
7
Run a bounded repair pass or smaller functional fallback.
8
Create Atlas collections, seed data, and search indexes.
9
Publish and smoke-check the hosted app through the participant path.
10
Update live event state and send tracking and ready emails.
END-TO-END TESTING
The biggest improvement to iteration speed.
More than rendered HTML
The production runner follows access, submission, planning, approval, building, Atlas persistence, publication, the generated app, and the public gallery. It verifies that the app completes a real write.
Repeat the same path
Creating the end-to-end test harness rapidly improved iteration and hardening speed. The team could rerun the same production-like journey after each change instead of manually rediscovering integration failures.
Diagnostics become tests
The runner preserves diagnostics, turns failures into regression tests, and makes repeatability a product requirement across service boundaries.
BUILD ASSISTANT