Inspiration
Creators and teams waste hours doing repeatable work for every single video/post: brainstorming hooks, researching facts, writing scripts, crafting SEO titles/tags, generating thumbnail prompts, and then exporting everything into a usable “bundle”.
The bigger pain: LLMs are unreliable in production. Even when the idea is great, the output often breaks your UI because of invalid JSON (extra text, missing braces, timestamps like 7:30, raw newlines, or Python-style quotes). We built Creative Production Engine to solve both problems:
Turn one topic into a publish-ready content kit
Make it robust enough for real product demos, not just prompts
What it does
Input: Topic + tone + language Output: A complete “Production Kit” that includes:
5 Hooks (curiosity openers)
5 High-CTR Titles
SEO Description
10 Tags
Thumbnail Text + Thumbnail Image Prompt
5 Shorts (each with title + 25–45 sec script)
Generate a script
Export the full kit as a plain-text “ready-to-use” bundle
Key product feature:
Regenerate only one section (e.g., just Titles / Hooks / Shorts / Script) without rebuilding everything.
How we built it
Backend (Django + DRF)
Built REST APIs to generate a full kit and regenerate individual sections.
Integrated local Ollama (no paid APIs) to generate structured outputs.
Implemented a Progressive JSON Reliability Pipeline so the system remains stable even when the model output is malformed:
Extract first JSON object (remove junk before/after)
Balance missing {} / []
Quote invalid time tokens like 7:30 → "7:30"
Convert Python-style 'single quotes' → valid JSON "double quotes"
Fix multiline JSON strings by escaping newlines as \n
Final fallback: parse as Python literal using ast.literal_eval
Frontend (Next.js)
Clean UI for:
Generating a kit from a topic
Viewing sections (Hooks, Titles, Description, Tags, Thumbnail, Shorts, Script)
Per-section regeneration buttons with loading spinners
Export download button
A “Recent Kits” sidebar for fast demo navigation
Storage / History
Saved recent kits to SQLite DB (ProductionKit model) to support a real workflow.
Keeps the last N kits to stay demo-friendly and fast.
Tech stack used (high level):
Django, Django REST Framework, CORS, Requests, local Ollama
Next.js (React) UI
Optional Postgres support via psycopg2 (works with SQLite too)
Challenges we ran into
LLM output breaks apps
Even one invalid token can crash the entire UI.
We solved it with a layered repair + parsing strategy and returned a safe “kit shape” even on failure.
Multiline content in JSON
Scripts/descriptions often include newlines which can break JSON.
We encoded and repaired strings safely (escaping \n).
Per-section iteration without data drift
Regenerating titles should still match topic/tone and stay consistent with the kit.
We passed context (existing hooks/titles) and added strict per-section output rules.
Hackathon demo reliability
A demo can fail if the model fails once.
Our backend never hard-crashes the UI; it returns debug info (raw, fixed, hint) so the demo still works.
Accomplishments that we're proud of
Built a working end-to-end “content factory”: topic → kit → export bundle
Added a production-grade reliability layer for JSON (the biggest real-world blocker)
Implemented section-level regeneration (fast iteration like a real creator tool)
Added history so judges can see multiple runs and real usage patterns
No paid AI APIs: runs locally with Ollama
What we learned
The difference between a “cool prompt” and a real product is reliability.
LLMs are powerful, but you must build defensive layers:
extraction, repair, validation, type-safety, and graceful fallbacks
UX matters: per-section regen + history makes it feel like a real tool, not a one-off generator.
What's next for Creative Production Engine
Distribution & Analytics (Future Phase)
Publish/schedule outputs (YouTube, Instagram, blogs)
Track CTR, retention, engagement
Auto-compare titles/thumbnails and recommend improvements
Create a “Learning Loop” where analytics updates:
tone rules
script structure
title patterns
thumbnail style prompts
Scaling product direction
Multi-domain config presets (education, finance, storytelling, SaaS, news)
Team workflows (templates, approvals, reusable brand voice)
Optional on-device fine-tuning / domain adaptation for consistent style
Log in or sign up for Devpost to join the conversation.