PASE — Population Agent Simulation Engine
Your Clients Are Making $50M Decisions on 2,000 Opinions
A CPG brand testing a campaign with Gen Z women in the Midwest: $150K, 8-week traditional study vs. PASE: 10,000 agents, full results with demographics, strategic insights, and audio podcast — under $8, in 3 minutes.
“That’s not a prototype demo. That’s a production system running right now on Google Cloud.”
Laurent Madeleni — AKQA | April 2026 | Built on Google Cloud Platform | Powered by JEPA + Gemini
The Platform
This Is Live. Right Now.
Four distinct workflows cover every research scenario — from replaying existing studies in new markets to running full what-if simulations with economic and media levers. Each one runs on the same production infrastructure, right now.
REPLAY A SURVEY
“We fielded this Pew study last quarter. What would it look like if we ran it in Miami-Dade right now?”
Select survey → Pick location → Define audience → Configure run → View results.
GENERATE NEW STUDY
“The client just called — they want to know how parents feel about screen time by tomorrow.”
Build from scratch, Likert/MC/binary/open-ended/ranking. Blank page to results in under 10 minutes.
TEST A CONCEPT
“We’ve got three campaign directions. Which one resonates with affluent suburban moms?”
A/B/C variants, auto-generated diagnostic battery: awareness, appeal, relevance, purchase intent, differentiation.
WHAT-IF SCENARIOS
“If unemployment spikes to 8% and media turns negative, how does that change brand perception?”
Economic levers (Unemployment, Gas Price, CPI, Consumer Sentiment, Fed Funds, Mortgage, VIX) + Media levers (Tone, Goldstein Score, Conflict %, Protest %, Event Volume).
The Payoff
You Don’t Get a Spreadsheet. You Get a Ready-to-Present Story.
Sample Distribution
Agentic Insight Pipeline
OBSERVE
AI reads every result, finds patterns a human would need hours to surface.
STRATEGISE
Writes the “so what” — strategic implications, opportunities, recommended actions.
VALIDATE
Self-checks every claim against actual data. No hallucinated stats.
COMPOSE
Designs scrollytelling HTML report, branded, responsive, with charts and pull quotes.
PODCAST
Narrates the report as audio briefing. MP3 download or in-app player.
PUBLISH
Hosted URL, version history, compare insights across simulations over time.
AGENT INTERVIEWS
“Talk to” individual agents. Q&A with a 28-year-old Black woman in Atlanta about why she rejected your campaign. With audio.
COMPARE EVERYTHING
Same study, different audiences/scenarios/times. Side-by-side with delta heatmaps.
Proof Points
Real Studies. Real Insights.
Every study below was run on the live PASE platform. These are not mock-ups or projections — they are actual simulation results with real strategic findings that would cost hundreds of thousands of dollars and months of time through traditional methods.
Starbucks Brand Health
Key Finding
5.1% rate likelihood to recommend as “0” despite product satisfaction. Non-dairy upcharge identified as most disliked policy (37.5% strongly disagree).
Airbnb Trust & Safety
Key Finding
57% worry about hidden cameras. “Unreliable” (35.6%) outranks “Authentic” (14.5%). 90%+ familiarity yet severe trust erosion.
Pepsi Brand Perception
Key Finding
91%+ familiarity but 25.7% lack emotional connection. Zero-sugar = major growth opportunity (85.8% interest). Coca-Cola remains dominant challenge.
Starbucks App Deep Dive
Key Finding
42% “Very dissatisfied” with core functionality, yet 40% rate app superior to competitors. 42.8% would visit less without it — dependency vs. frustration paradox.
The Science
Trained on 23,126 Real Americans
REAL PEOPLE
23,126 respondents, 63 waves Pew American Trends Panel (Sep 2021–Nov 2024). 483K response rows.
REAL DYNAMICS
Neural dynamics engine (JEPA) learned how real people changed over time. Ingests Federal Reserve data + 48M news events from GDELT.
CENSUS GROUNDING
Anchored to 84,414 US Census tracts, 240+ demographic variables.
GOOGLE SEARCH GROUNDING
Every question sent to Gemini + Google Search for real-world baseline before JEPA dynamics adjusts.
Most synthetic research tools generate from a frozen model. PASE queries the live internet for every question, for every segment, before generating a single answer.
Validation Metrics
better than naive baseline
attitude variables outperformed
domains validated (politics, brand, tech, health, social)
countries (US + UK)
cross-country transfer
“60+ experiments. We broke it every way imaginable. Then we ran it on British election data without retraining and it still worked.”
The Economics
What Used to Cost $150K Now Costs $8
The economics of synthetic population research fundamentally change what’s possible. When the cost of asking a question drops by four orders of magnitude, every strategic decision can start with data.
The cost of curiosity drops to nearly zero. Every pitch can start with data. Every strategy can be tested before presented.
LLM Tier Costs
Open-ended
Gemini 2.5 Pro
~$15–30/1K agents
Structured
Gemini 2.5 Flash
~$3–8/1K agents
Binary/MC
Gemma 3 (self-hosted)
~$0.50–2/1K agents
The Frontier
API-First. Edge-Ready. And About to Get Much Faster.
PASE is architected for integration at every level — from direct API calls to autonomous AI agents running surveys on your behalf. The progression from server to edge means costs approach zero for structured questions.
API-FIRST ARCHITECTURE
Every feature is an endpoint. REST API with endpoints listed. Python + TypeScript SDKs coming.
MODEL CONTEXT PROTOCOL (MCP)
Let any AI assistant run surveys. MCP server for PASE. Strategist says “Run brand health check on Starbucks among Gen Z women in Florida” → agent calls PASE.
WPP OPEN INTEGRATION
PASE as MCP server in WPP Open’s agent framework. Tools: create_survey, run_simulation, get_results, generate_insight. Research as a tool call.
CLIENT CLOUD DEPLOYMENT
License model. JEPA model ~400K params. Containerised, deployed to client’s GCP/AWS/Azure. Client pays compute, WPP licenses the model. Margin approaches 100%.
Edge Inference
~2MB ONNX artifact, runs in browser (WebAssembly) or phone. Sub-second, zero-cost for structured questions.
The Roadmap
The Next 90 Days
From working platform to client-ready product.
Already Delivered
JEPA dynamics engine (34.3× baseline)
Agentic insight pipeline (6-step)
Census-grounded population synthesis (84,414 tracts)
Podcast audio generation
4 survey workflows
Agent interview synthesis
Google Search grounding via Gemini
Real-time economic & media lever injection
Interactive scrollytelling HTML reports
Delta heatmap comparison engine
Hosted URL publishing with version history
Sprint Plan
Sprint 1
Apr 14–25
Scenario Engine & Sentiment
Sprint 2
Apr 28 – May 9
Longitudinal & Comparison
Sprint 3
May 12–23
Client Readiness & Vertical Customisation
Sprint 4
May 26 – Jun 6
UK Expansion & Scale
Sprint 5
Jun 9–20
Pilot Launch & Enterprise
July 2026 Target
July 2026: Production-grade platform with UK market coverage, enterprise client pilots, and full API + MCP integration for WPP Open.
The Ask
This Works. Let’s Make It Ours.
PASE is built. It runs. It produces results that would cost hundreds of thousands of dollars through traditional methods. What it needs now is momentum, pilots, and connections to the data that will make it unstoppable.
GIVE US MOMENTUM
WPP Media’s Momentum survey = training signal for global expansion. 70+ countries, $0 refresh cost, 24/7 population availability. “Not a replacement — an amplifier.”
PILOT IT
Pick 2–3 client briefs. Run alongside traditional research. CPG, Auto, Financial Services.
CONNECT IT
GroupM media intelligence, Kantar panels, Wunderman Thompson CRM. More data = more markets = more clients.
GPU cost is <$500/month. Infrastructure is built.
“The question isn’t whether AI will transform research. It’s whether we’re the ones who do it.”
— Laurent Madeleni — AKQA