Inspiration
Big Bytes is a research and editorial workflow that turns a broad search into a small set of stories worth reviewing. The original pipeline worked, but it was hard to understand what was happening during a run or tune the workflow without editing configuration. Big Bytes Factory turns that invisible pipeline into a live, playable factory floor.
What it does
Big Bytes Factory is a Factorio-inspired visual workspace for web research and editorial triage.
- Exa discovers candidate stories.
- A scanner and dedupe stage clean the result set.
- An optional Firecrawl courier fetches thin source pages.
- GPT-5.6 reviews the evidence and assigns Queue, Maybe, or Skip.
- Accepted items leave the editor with evidence, a canonical source, and a copy-ready draft.
- Convex streams runs, events, items, and factory layout changes live to the UI.
The factory layout is functional, not decorative. Builders can place and move machines, paint belts, change courier capacity, and use the resulting layout as the run configuration. A run only starts when the required machines are connected in a valid chain.
How we built it
The app uses Next.js 16, React 19, TypeScript, the Canvas 2D API, and Convex. The pipeline runs on the Next.js server and writes its progress to Convex, while the React client subscribes to live updates without polling.
Provider adapters integrate Exa, Firecrawl, and the OpenAI Responses API. A deterministic mock-provider mode makes the full animated experience testable without paid calls. The project also includes Node-based pipeline and factory checks plus browser acceptance tooling.
A validated real-provider run completed with 40 unique Exa candidates from 34 domains and eight GPT-5.6 editorial records. The persisted OpenAI receipt identified the model as gpt-5.6-sol, and the Convex run finished with no error. Firecrawl remained available as a fallback but was not needed for that run.
How Codex and GPT-5.6 were used
Codex was the primary engineering partner for the rewrite. It traced the original Python pipeline, designed the Next.js/Convex architecture, implemented the canvas factory and provider pipeline, repaired integration issues, wrote the smallest useful automated checks, and drove production-build browser verification.
Codex also caught an important validation failure: an early credential-looking run had actually used mock GPT output. It traced the stale server path, restarted the app with mock mode explicitly disabled, and verified the real run from provider metadata, saved receipts, source URLs, and Convex state instead of trusting a health check or HTTP 202 response.
GPT-5.6 is also part of the product itself. It receives structured source evidence and returns editorial verdicts, reasoning, and copy-ready drafts for each selected story.
Challenges
The hardest part was keeping a playful visual system tied to real pipeline state. Layout changes needed to be editable and persistent, while active runs had to remain deterministic. Each run therefore snapshots its layout before execution.
The second challenge was proving that a provider run was genuinely live. Readiness endpoints were insufficient, so the verification flow checks model metadata, persisted receipts, non-mock source URLs, and the final Convex record.
Accomplishments
- A coherent pixel-art product experience rather than a pipeline dashboard.
- Live Convex subscriptions with no client polling.
- Factory topology that directly configures and validates execution.
- Deterministic zero-cost demo runs.
- A verified Exa-to-GPT-5.6 run with persisted evidence.
- Automated pipeline, layout, and browser acceptance checks.
What we learned
Agent-assisted development is strongest when the agent is asked to verify outcomes, not just generate code. Codex accelerated the rewrite, but its biggest contribution was tracing the full system and refusing to treat “provider ready” as proof that a real model call had completed.
What's next
The next useful steps are durable background execution, authentication, richer run history, media inspection, and publishing/export workflows.
Built With
- canvas-api
- codex
- convex
- exa
- firecrawl
- gpt-5.6
- next.js
- node.js
- openai-api
- react
- typescript