yogen - 予言
yogen - 予言
yogen
The Question
What if you could pressure-test a startup idea against 500 simulated minds before spending a single penny? Not a chatbot. Not a survey. A genuine debate. Experts with identities, biases, and the ability to change their minds mid-argument.
What if you could pressure-test a startup idea against 500 simulated minds before spending a single penny? Not a chatbot. Not a survey. A genuine debate. Experts with identities, biases, and the ability to change their minds mid-argument.
What if you could pressure-test a startup idea against 500 simulated minds before spending a single penny? Not a chatbot. Not a survey. A genuine debate. Experts with identities, biases, and the ability to change their minds mid-argument.
Solution
Yogen is that. A swarm of 500 AI agents, 12 named experts plus 488 background agents — debate your idea across 6 rounds and return a verdict with a full report.
Yogen is that. A swarm of 500 AI agents, 12 named experts plus 488 background agents — debate your idea across 6 rounds and return a verdict with a full report.
Yogen is that. A swarm of 500 AI agents, 12 named experts plus 488 background agents — debate your idea across 6 rounds and return a verdict with a full report.
Outcome
Built, shipped, live - 2 weeks from idea to deployed product.
Built, shipped, live - 2 weeks from idea to deployed product.
Built, shipped, live - 2 weeks from idea to deployed product.
Year
2026
Role
Founder — product, engineering, design
Stack
Next.js · Claude Sonnet · SSE · Railway
Status
How it works
How it works
You submit an idea. Claude generates 12 expert agents tailored to your domain - each with a name, role, stance, and bias. They debate across 6 rounds. After each round, stances update. Background agents follow dive in. The canvas updates live. At the end, a verdict, full report and next steps are generated.
You submit an idea. Claude generates 12 expert agents tailored to your domain - each with a name, role, stance, and bias. They debate across 6 rounds. After each round, stances update. Background agents follow dive in. The canvas updates live. At the end, a verdict, full report and next steps are generated.
You submit an idea. Claude generates 12 expert agents tailored to your domain - each with a name, role, stance, and bias. They debate across 6 rounds. After each round, stances update. Background agents follow dive in. The canvas updates live. At the end, a verdict, full report and next steps are generated.
Meet the cast
Meet the cast
Each expert is generated fresh per simulation — name, background, bias, and stance assigned by Claude based on the idea's domain. They are not templates. Each one of the 12 are distinct personalities.
Each expert is generated fresh per simulation — name, background, bias, and stance assigned by Claude based on the idea's domain. They are not templates. Each one of the 12 are distinct personalities.
Each expert is generated fresh per simulation — name, background, bias, and stance assigned by Claude based on the idea's domain. They are not templates. Each one of the 12 are distinct personalities.
Maya Chen
Series B Investor · Consumer Tech
SCEPTICAL
Dr. Priya Nair
Regulatory Scientist · Life Sciences
NEGATIVE
David Osei
Brand Strategist · D2C
POSITIVE
Aiko Tanaka
Product Lead · B2B SaaS
POSITIVE
Marcus Webb
CFO · Scale-up
SCEPTICAL
Sofia Reyes
UX Researcher · Enterprise
POSITIVE
James Okafor
Growth Lead · Marketplace
SCEPTICAL
Nia Mensah
Impact Investor · Climate
POSITIVE
Tom Hargreaves
CTO · Infrastructure
NEGATIVE
Layla Hassan
CMO · Consumer Brand
POSITIVE
Ravi Patel
Ops Director · Logistics
SCEPTICAL
Clare Dunne
Legal · IP & Contracts
NEGATIVE
The Centrepiece — The Swarm Canvas
The Centrepiece — The Swarm Canvas
The Centrepiece
The canvas is the centrepiece. 500 agents rendered as dots, grouped into three clusters — Positive, Sceptical, Negative. They move. They drift. They flip. You watch the crowd shift in real time as the experts debate.
The canvas is the centrepiece. 500 agents rendered as dots, grouped into three clusters — Positive, Sceptical, Negative. They move. They drift. They flip. You watch the crowd shift in real time as the experts debate.
The canvas is the centrepiece. 500 agents rendered as dots, grouped into three clusters — Positive, Sceptical, Negative. They move. They drift. They flip. You watch the crowd shift in real time as the experts debate.
A — Clusters
Three live clusters. Dots are background agents. Size of cluster reflects proportion. They move toward the cluster that matches their current stance.
A — Clusters
Three live clusters. Dots are background agents. Size of cluster reflects proportion. They move toward the cluster that matches their current stance.
A — Clusters
Three live clusters. Dots are background agents. Size of cluster reflects proportion. They move toward the cluster that matches their current stance.
A — Clusters
Three live clusters. Dots are background agents. Size of cluster reflects proportion. They move toward the cluster that matches their current stance.
B — Name chips
The 12 expert agents surface as named chips above their cluster. A filled black chip means they just spoke. The chip pulses on their turn.
B — Name chips
The 12 expert agents surface as named chips above their cluster. A filled black chip means they just spoke. The chip pulses on their turn.
B — Name chips
The 12 expert agents surface as named chips above their cluster. A filled black chip means they just spoke. The chip pulses on their turn.
B — Name chips
The 12 expert agents surface as named chips above their cluster. A filled black chip means they just spoke. The chip pulses on their turn.
C — Flip banner
When an expert changes stance — positive to sceptical, sceptical to negative — a banner fires at the bottom. The crowd watches experts for direction.
C — Flip banner
When an expert changes stance — positive to sceptical, sceptical to negative — a banner fires at the bottom. The crowd watches experts for direction.
C — Flip banner
When an expert changes stance — positive to sceptical, sceptical to negative — a banner fires at the bottom. The crowd watches experts for direction.
C — Flip banner
When an expert changes stance — positive to sceptical, sceptical to negative — a banner fires at the bottom. The crowd watches experts for direction.
Under the Hood — The Generation Pipeline
Under the Hood — The Generation Pipeline
Under the Hood
Before the debate starts, Claude generates the entire expert panel from scratch, tailored to the idea's domain. Early versions had agents who were generic and interchangeable, drifting toward the same opinion by round three. This took many iterations to fix.
Before the debate starts, Claude generates the entire expert panel from scratch, tailored to the idea's domain. Early versions had agents who were generic and interchangeable, drifting toward the same opinion by round three. This took many iterations to fix.
Before the debate starts, Claude generates the entire expert panel from scratch, tailored to the idea's domain. Early versions had agents who were generic and interchangeable, drifting toward the same opinion by round three. This took many iterations to fix.
STEP 01
Archetype Assignment
8 archetypes — investor, sceptic, regulator, user, competitor, media, operator, advocate. Each has a default starting stance: investors start sceptical, advocates start positive, competitors start negative. This seeds the debate with real tension.
STEP 02
Persona Generation
One Claude call generates all 12 personas. Haiku model, JSON schema enforced, ~2s total.
name: “Maya Chen”
role: “Series B investor, consumer goods”
bio: “Led 14 investments in DTC brands…”
bias: “Demands verifiable unit economics”
style: “Direct. Numbers before narrative.”
name: “Maya Chen”
role: “Series B investor”
bio: “Led 14 investments in DTC”
bias: “Demands verifiable unit”
style: “Direct. Numbers only.”
STEP 03
Knowledge Slicing
Full seed doc (~3000 chars) is sliced per agent — only the facts relevant to that agent’s domain get passed in. The regulator gets compliance detail, the investor gets market size. Stops agents getting confused by irrelevant context.
STEP 04
Prompt Caching
The stable part of each agent’s system prompt (persona + domain knowledge) is cached after the first call; only the variable suffix (round, peer statements) is fresh. Cuts token cost by ~60% across 9 agents × 6 rounds.
Going Inside - From Crowd to individual node
Going Inside - From Crowd to individual node
Going Inside -
The swarm is three views in one. The macro view shows the crowd. The mid view shows clusters separating. The micro view — when you hover or click — shows individual agents and what they’re thinking. Building this meant solving a canvas rendering problem: 500 moving dots at 60fps without jank.
The swarm is three views in one. The macro view shows the crowd. The mid view shows clusters separating. The micro view — when you hover or click — shows individual agents and what they’re thinking. Building this meant solving a canvas rendering problem: 500 moving dots at 60fps without jank.
The swarm is three views in one. The macro view shows the crowd. The mid view shows clusters separating. The micro view — when you hover or click — shows individual agents and what they’re thinking.
VIEW 01 — MACRO
38%
POS
44%
SKEP
18%
NEG
Macro. Full swarm visible. Three clusters. Depth-of-field: foreground dots are larger and opaque, background dots smaller and transparent. Creates a real sense of crowd depth.
Macro. Full swarm visible. Three clusters.
Macro. Full swarm visible. Three clusters.
Depth-of-field: foreground dots are larger and
opaque, background dots smaller and transparent.
Creates a real sense of crowd depth.
Macro. Full swarm visible. Three clusters. Depth-of-field: foreground dots are larger and opaque, background dots smaller and transparent. Creates a real sense of crowd depth.
VIEW 02 — CASCADE EVENT
Cascade. High-influence agent flips. A tracer line fires from their position toward the cluster they’re joining. Nearby agents with high volatility may follow within the same round.
Cascade. High-influence agent flips.
Cascade. High-influence agent flips. A tracer line
fires from their position toward the cluster they’re
joining. Nearby agents with high volatility may
follow within the same round.
Cascade. High-influence agent flips. A tracer line fires from their position toward the cluster they’re joining. Nearby agents with high volatility may follow within the same round.
VIEW 03 — MICRO / HOVER
bg_284
archetype: investor
stance: sceptical
volatility: 0.72
Micro. Hover on any dot. A soft ring appears. For background agents you see their archetype, stance, and volatility score. For experts, you get the full tooltip with their last statement.
Micro. Hover on any dot. A soft ring appears.
Micro. Hover on any dot. A soft ring appears.
For background agents you see their archetype,
stance, and volatility score. For experts, you get
the full tooltip with their last statement.
Micro. Hover on any dot. A soft ring appears. For background agents you see their archetype, stance, and volatility score. For experts, you get the full tooltip with their last statement.
HOW THE CANVAS WAS BUILT
HOW THE CANVAS WAS BUILT
Rendering
HTML Canvas 2D. Each frame: clear rect, re-draw all 500 dots. requestAnimationFrame loop. No WebGL — didn’t need it for this count.
Depth of field
Each dot has a z-value. z determines radius (2–5px) and opacity (0.3–1.0). Sorting by z on each frame gives the depth illusion.
Movement
Dots drift with perlin-like noise toward their cluster target. Velocity is capped so the swarm doesn’t spin. Bounce margin keeps them on screen.
Hit detection
On mousemove, iterate all 500 agents, calc distance to cursor. Nearest within 12px gets the tooltip. O(n) per frame — fast enough.
Full Sim
Full Sim
While the swarm animates, the debate feed streams alongside it. Every agent turn arrives word-by-word via SSE — a 4-line preview by default, full text on toggle. When an expert flips stance, the feed shows the change inline. You can also speak as the founder, injecting a message mid-debate that agents may reference, challenge, or ignore depending on their bias
While the swarm animates, the debate feed streams alongside it. Every agent turn arrives word-by-word via SSE — a 4-line preview by default, full text on toggle. When an expert flips stance, the feed shows the change inline. You can also speak as the founder, injecting a message mid-debate that agents may reference, challenge, or ignore depending on their bias
While the swarm animates, the debate feed streams alongside it. Every agent turn arrives word-by-word via SSE — a 4-line preview by default, full text on toggle. When an expert flips stance, the feed shows the change inline. You can also speak as the founder, injecting a message mid-debate that agents may reference, challenge, or ignore depending on their bias
The report
The report
The debate doesn't end when the canvas stops moving. Every simulation closes with a generated report — PDF-ready, built from the actual transcript: the verdict, the requirement checklist, a round-by-round stance breakdown for all 12 experts, and their most decisive quotes. It's the thing you actually take away and act on.
The debate doesn't end when the canvas stops moving. Every simulation closes with a generated report — PDF-ready, built from the actual transcript: the verdict, the requirement checklist, a round-by-round stance breakdown for all 12 experts, and their most decisive quotes. It's the thing you actually take away and act on.
The debate doesn't end when the canvas stops moving. Every simulation closes with a generated report — PDF-ready, built from the actual transcript: the verdict, the requirement checklist, a round-by-round stance breakdown for all 12 experts, and their most decisive quotes. It's the thing you actually take away and act on.
The Outcome
The Outcome
(2014-26)
Let’s talk.
Tell me about your project-whether it’s a app, website, product creation or a build with Ai.
Quick response.
If you’re ready to create and collaborate, I’d love to hear from you.
Lead Product Designer
Moodflo
Michael Ruocco
Let’s talk.
Tell me about your project-whether it’s a app, website, product creation or a build with Ai.
Quick response.
If you’re ready to create and collaborate, I’d love to hear from you.
Lead Product Designer
Moodflo
Michael Ruocco
Let’s talk.
Tell me about your project-whether it’s a app, website, product creation or a build with Ai.
Quick response.
If you’re ready to create and collaborate, I’d love to hear from you.
Lead Product Designer
Moodflo
Michael Ruocco