Multi-Agent Editorial System

🏆 Winner of BSEC 2026 Hackathon (Brno University of Technology)

1. Project Overview

We built this system in a 12-hour sprint with a clear goal: to create a live-demo-ready workflow that helps creators decide what to film, where to film it, and how to publish it across platforms instantly.

The result is a functional editorial workflow that takes a short brief and produces high-quality outputs for:

• YouTube (video script)
• Instagram (post)
• TikTok (reel)

---

2) Contributors

Sorted alphabetically

Contributor	GitHub
LufyCZ	LufyCZ
olexamatej	olexamatej
Padi42	Padi42
TheRamsay	TheRamsay

---

3) Assignment (short)

Based on important_files/Multi-Agentná AI Redakcia pre správu obsahu.pdf, the goal was to design and implement a working prototype of an editorial content generator for YouTube, Instagram, and TikTok.

Key constraints from the assignment:

• architecture and logic matter more than tooling choice,
• AI should be used efficiently (not for everything),
• clear multi-agent responsibilities in the editorial process:
  • Author: content proposal,
  • Editor: platform adaptation,
  • QA: final quality decision.

Required functionality:

• 2 required use-cases:
  1. generate topic options,
  2. take a short brief and generate outputs for all platforms;
• required brief fields: tema, cil, cilova_skupina, ton, hlavni_myslenka, cta;
• required outputs for all 3 platforms;
• use of creator history + web trend enrichment;
• structured QA approve/revise decision;
• token/cost optimization and minimal unnecessary AI calls;
• end-to-end completion in a few minutes;
• frontend showing complete workflow status;
• automatic email send before judging.

---

4) Our solution

A) Kamil dashboard agent + tool calling

Kamil is the central dashboard helper (inspired by Jarvis):

• understands user intent in plain language,
• calls tools and routes requests into research/generation flows,
• can be used both in the web dashboard and as a Telegram bot.

_{Kamil chat with tool-calling in action; user asks, Kamil decides whether to trigger research/workflows.}

_{Research results are returned in a readable assistant format, ready for immediate post creation.}

B) Research + anti-hallucination pipeline

Our product intent was practical: help an influencer quickly discover what to show while traveling in a specific city (places, hooks, trends, and content angles) and turn that into platform-ready content.

Our research flow is designed to reduce hallucinations:

1. fetches best matching creator-history examples,
2. runs trends + location search in parallel,
3. verifies locations in a second pass,
4. geocodes and returns structured context,
5. forwards validated context into downstream generation.

_{The anti-hallucination flow combines history retrieval, web search, verification, and structured output.}

_{Topic-related places are visualized on a map so creators can quickly plan what to shoot in that city.}

_{Trend signals and content angles are surfaced as practical ideas, not generic brainstorming.}

_{One-click transition from research to a post draft workflow.}

C) Swarm democracy (4-agent discussion)

We use a democracy-style swarm with 4 role-specialized AI personalities.

In the current implementation, swarm runs on one strong base model (Gemini 2.5 Flash) but with different editorial personalities and responsibilities:

• Zoey (Gen-Z trend scout): focuses on hooks, viral formats, and platform-native energy.
• Marcus (Producer): focuses on structure, pacing, retention, and execution quality.
• Dr. Chen (Analyst): focuses on originality, depth, and non-obvious angles.
• Leo (Reviewer/QA): challenges ideas and scores platform fit, clarity, and performance potential.

Each round has generation/revision + peer rating, then score aggregation picks the strongest direction.

This creates better ideas than a single-shot prompt.

_{Multi-agent voting/revision loop where multiple perspectives compete before final selection.}

_{Project detail view showing swarm status, generated outputs, and next actions in one place.}

D) Project creation + editorial workflow UI

The app supports full project lifecycle:

• create projects,
• create/manage posts,
• attach media and context,
• track generation/QA status per platform,
• inspect final outputs.

_{Global project list and progress overview.}

_{Simple project creation flow to start a new campaign quickly.}

_{Per-project post management with clear state tracking.}

_{Brief, media, and generated content are kept together for fast editorial iteration.}

E) Image understanding ("what is on the image")

The system does automatic image understanding to ground output in visual evidence:

• vision model describes the image content,
• semantic text + vector embeddings are stored,
• relevant images are ranked by similarity and injected into prompts.

This helps avoid hallucinating visuals and improves relevance.

F) Thumbnail generation + video maker

Beyond text generation, we also support:

• automatic YouTube thumbnail generation,
• Remotion-based video making pipeline (script/speech/video composition).

_{Preview of automated video generation flow (script → speech → composited short-form video).}

G) QA + email to judges

Each platform output receives a structured QA decision (approve / revise) with reason and recommended edits.

The final output can be sent by email for judging.

_{Final output packaged into email-ready format for judges/comparison.}

---

5) Other things

Tech stack

• Next.js (App Router), React, TypeScript
• Gemini/Vertex AI (@google/genai, @ai-sdk/google-vertex)
• Drizzle ORM + Turso/libSQL
• UploadThing
• Remotion
• Nodemailer + React Email