We ran MiroFish on a real brand campaign. 105 AI agents. 6 minutes. Here's what happened.

MiroFish hit GitHub #1 trending last week (28K stars). Everyone's asking: what can you actually build with it?

We built Campaign Pressure Test™ — a brand safety simulation for marketing teams. Here's a real run on a Singapore fashion brand (Nimisski) launching a spring campaign on Xiaohongshu.

The Problem

Brands spend $50K–$500K on campaigns that flop because no one tested whether the audience would actually respond. Traditional focus groups take 2–4 weeks and cost $15K+. By the time you know it won't work, you've already bought the media.

What We Built on Top of MiroFish

Standard MiroFish takes any seed document and spawns AI agents. We added one proprietary layer: Style Genome™ injection.

Before simulation, we embed a 768-dimensional brand memory vector into the seed document. This makes every agent brand-aware — they know Nimisski's price positioning ($189–449 SGD), aesthetic language (Korean-minimal), and competitive context before the simulation starts.

The result: agents don't just react to "a fashion campaign." They react to this brand's campaign, the way this brand's target consumers would.

Real Results (345 seconds, not simulated)

Brief fed to MiroFish:

"Spring Awakening campaign. Platform: Xiaohongshu. Market: Singapore. Korean-minimal fashion. Target: 25–35 urban women."

Output:

• 7 posts generated organically by agents
• 17 comments, 105 total actions
• Brand Safety Score: 9.2/10 ✅
• Engagement rate: 40% (likes ÷ total actions)

The competitor risk it caught — unprompted:

Agents independently introduced Pomelo with the message "wide range of trendy and affordable options." Zero instructions to mention competitors. The simulation surfaced that Nimisski's $189–449 price range would face organic competitive framing from Pomelo unless campaign copy proactively justified the premium.

That's signal you can't get from a survey.

The Technical Stack

# 9-step MiroFish pipeline
1. ontology/generate    # Build knowledge graph from seed
2. graph/build         # GraphRAG construction
3. graph/task poll     # Wait for graph ready
4. simulation/create   # Initialize sim with agent count
5. simulation/prepare  # Inject Style Genome seed here
6. prepare/status poll # Wait for agents to load
7. simulation/start    # Launch the swarm
8. run-status poll     # Monitor until complete
9. posts + comments + agent_stats  # Extract results

Key gotcha: Use gemini-2.0-flash, NOT gemini-2.5-flash. The thinking tokens in 2.5 cause 500 errors on MiroFish's simulation endpoint.

Brand Safety Score formula:

score = (positive_pct * 5.0) + \
        ((1 - negative_pct) * 3.0) + \
        ((1 - safety_flag_pct) * 2.0) - penalty

Why This Matters Now

MiroFish is still a raw engine — Docker + API key setup, no UI, no reporting. Most devs who clone it give up before getting value.

We wrapped it into a one-command CLI:

python run_pressure_test.py \
  --brief "Your campaign brief here" \
  --platform XHS \
  --industry fashion \
  --market Singapore

6 minutes → HTML/PDF report → Brand Safety Score.

The Competitive Moat

Every US synthetic research platform (Aaru $1B, Simile $100M) uses generic AI personas. None of them have brand memory. Our Style Genome™ vectors are trained on real campaign data from CanMarket's existing client base — Haleon, Starbucks, ByteDance, McCann.

An agent that knows your brand's DNA produces fundamentally different simulation outputs than a generic consumer persona.

Try It

• $499/simulation — DM for Stripe link or free demo
• Enterprise Monthly Guard: $2,500/month
• Free demo for your next campaign: comment below or DM

The Starbucks Harry Potter APAC campaign launches in 5 days. We're running a simulation tonight.

CanMarket is a Brand Operating System. "Run performance marketing without breaking your brand." MWC 2025 Global Runner-up.