Methodology — Mirra Labs

Example: Singapore — mass-affluent wealth consolidation

Step 1 of 6

The question we set out to answer

Every study starts with a real commercial decision — framed with stakes, constraints, and the specific action someone needs to take.

Private Bank MAS Regulatory Frame Offshore wealth outflow signal

Which mass-affluent Singaporean segments would consolidate their wealth relationship with a single bank — if offered an integrated private-banking tier?

At stake

AUM capture

At stake

Cross-sell depth

At stake

Segment prioritization

We're not asking who might say yes. We're asking who would act, why, and what would stop them.

Example: Vietnam — urban households

Step 2 of 6

Designing the population, segments, and behavioral frame

Populations start from published, official sources. Segments are designed — not discovered — so that every finding can be traced back to a decision frame you recognize.

Calibrated against

Official national sources

The hard baseline — census, credit, regulatory filings.

Multilateral cross-checks

Independent views that triangulate the baseline.

Calibrated segment matrix

Urban Vietnamese households — income × life stage.

Young single

Young family

Established family

Empty nester

Affluent

38K HH

High-earner professionals

82K HH

Dual-career, no kids yet

124K HH

Peak-income households

71K HH

Pre-retirement wealth

Upper
income

142K HH

Urban single professionals

318K HH

Dual-income, 1 child

405K HH

School-age household

211K HH

Children out of home

Mid
income

486K HH

Early-career, living alone

812K HH

Saving for first home

954K HH

Education-focused spend

523K HH

Helping adult children

Lower-mid
income

672K HH

Informal / gig workers

1.04M HH

Dependent on one earner

1.18M HH

Multi-generational home

698K HH

Pension-dependent

1 The Validation Protocol — Check 1

Source Triangulation

Every population anchor is cross-checked against at least two independent sources before any agent is built.

Example: Indonesia — young urban professionals

Step 3 of 6

One segment. Thousands of distinct agents.

No two agents are alike — and the differences between them are where the strategic signal lives.

One segment

Young urban professional

Age 25–34 · Jakarta, Surabaya, Bandung · Monthly income Rp 8–20M

Thousandsof distinct agents generated inside this segment — each with a different life, a different decision calculus, a different story.

Faith-aligned banking

Raka · 28 · M

Jakarta · Copywriter · Rp 11M/mo

Behavioral traits

Avoids riba — BSI only, no credit card
Saving for umrah in 3–5 years

Digital-first investor

Reza · 32 · M

Jakarta · Product Manager · Rp 18M/mo

Behavioral traits

Livin' + Bibit + Pluang across 3 apps
ETF and crypto-curious

Heavy credit user

Andi · 31 · M

Surabaya · Software Eng · Rp 20M/mo

Behavioral traits

Two CCs, pays in full monthly
Regional travel, FX fee-sensitive

Household builder

Sari · 33 · F

Jakarta · Lawyer · Rp 16M/mo

Behavioral traits

Dual-income, saves jointly
Researching KPR for first home

Variable cashflow

Dimas · 29 · M

Jakarta · Freelance UX · Rp 8–22M/mo

Behavioral traits

Income swings 3x month-to-month
Holds 4-month emergency buffer

Cash-anchored

Putri · 32 · F

Surabaya · Teacher · Rp 9M/mo

Behavioral traits

BRI payroll, cash for daily spend
Low digital comfort, debt-averse

Lifestyle spender

Faiz · 27 · M

Bandung · Sales Exec · Rp 14M/mo

Behavioral traits

Uses BNPL for lifestyle purchases
Dining, concerts, new gadgets

Owner-operator

Bagus · 30 · M

Surabaya · Family shop · Rp 12–18M/mo

Behavioral traits

Mixes personal and business funds
Supplier credit, seasonal inventory

Global commuter

Intan · 28 · F

Jakarta · Consultant · Rp 22M/mo

Behavioral traits

Wise for FX, monthly client travel
Rate and fee sensitive

Pension-anchored

Budi · 33 · M

Bandung · PNS · Rp 10M/mo

Behavioral traits

BNI + Taspen, 20-year horizon
Strongly debt-averse

Shift-income saver

Ratna · 29 · F

Surabaya · Nurse · Rp 9M/mo

Behavioral traits

Irregular shift bonuses, cash-heavy
Gold saver to hedge inflation

Micro-entrepreneur

Yuni · 31 · F

Bandung · Warung · Rp 7–15M/mo

Behavioral traits

QRIS daily turnover, no biz account
Supplier credit for inventory

12 shown here — thousands generated per study. Each agent carries a distinct identity, economic reality, and decision calculus; variance is calibrated to real population distributions.

Without calibrated diversity

One segment treated as one response

The average collapses the real variation — you see a single answer where none exists.

With calibrated diversity

Multiple real behaviors, visible at once

Distinct clusters emerge — the signal you actually need to act on.

Example: Thailand — elder care product adoption

Step 4 of 6

Asking the right questions, the right way

Large language models answer some questions brilliantly and others poorly — the craft is knowing the difference. We decompose every research question into the sub-inquiries agents answer most reliably, then reassemble the answers into the strategic finding.

The research question

Will middle-income Thai families adopt at-home elder care services in the next 24 months?

Affordability signal

What would this cost relative to what you already spend caring for parents today?

Cultural frame

How does your family currently share elder-care responsibility across generations?

Trust & safety

What would have to be true for you to trust a stranger alone with your parents?

Peer influence

Who in your extended family or circle would need to try this first?

Alternative paths

What are you doing instead today — and how well is it actually working?

Trigger conditions

What event would force you to decide this within the next six months?

Reassembled →

A weighted finding: who adopts, at what price point, under what conditions — with the real blockers surfaced, not the stated ones.

We don't ask agents to predict the future. We ask them what they'd consider, what would stop them, and what would change their mind — then we assemble the answer.

2 The Validation Protocol — Check 2

Blind Model Review

A separate model audits the question design without access to the population or context, flagging any sub-inquiry an agent might answer unreliably.

Example: Malaysia — electric vehicle adoption

Step 5 of 6

Individual reasoning becomes collective pattern

The output isn't a vote count — it's a map of where conviction is highest, where it breaks down, and why.

Individual agent responses — thousands per study

↓

Clustering by reasoning similarity

↓

Urban Gen Z

Charging-access gap in condos and rentals — willingness is high but home-charging is structurally blocked.

Suburban families

Price elasticity kicks in sharply around the EV premium over ICE — narrow the gap and adoption follows.

Rural households

Range anxiety and thin rural charging networks are the binding constraint — infrastructure beats incentives.

Strategic signal map — Malaysia EV adoption segments

Segments plotted by barrier severity vs. willingness to adopt. Dot size reflects segment market value.

3 The Validation Protocol — Check 3

Behavioral Backtest

We backtest our agents against questions whose answers are already known, then carry that calibration into the questions where they aren't.

Step 6 of 6

Findings shaped for the people who have to act on them

Three deliverables, one underlying analysis — each shaped for the audience that will act on it.

Behavioral Segment Report

"Who are the people, and how do they respond?"

Mirra Labs Behavioral Segment Report — Indonesia

Context, methodology, and segment-by-segment findings — our primary written deliverable.

Visual Exhibits (Slide Deck)

"What are the key insights of the study?"

Client briefing deck — sharia framing insight

Client briefing deck — chart-heavy analytical slide

Strategic recommendation — decision-grade, action-ready.

Raw Data Set

"The full dataset behind the insights."

Underlying data pack — segment-level breakdowns and reasoning traces

The full depth behind every insight — slice it, pressure-test it, go deeper.

Every engagement ships a full report, a decision-grade briefing, and the underlying data pack. You don't just get the answer — you get the ability to see why.

4 The Validation Protocol — Check 4

Firewall Audit

Any reference data used to validate findings is sealed from the prediction pipeline — findings cannot be tuned to match known answers.

Validation Protocol — Final Trace

All checks passed

✓

Check 1Source Triangulation

✓

Check 2Blind Model Review

✓

Check 3Behavioral Backtest

✓

Check 4Firewall Audit

Academic grounding

Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Stanford University.

Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of One, Many: Using Language Models to Simulate Human Samples. Political Analysis.

Horton, J. J. (2023). Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus? NBER Working Paper.

Dillion, D., Tandon, N., Gu, Y., & Gray, K. (2023). Can AI language models replace human participants? Trends in Cognitive Sciences.

How we build behavioral intelligence

The question we set out to answer

Designing the population, segments, and behavioral frame

One segment. Thousands of distinct agents.

Asking the right questions, the right way

Individual reasoning becomes collective pattern

Findings shaped for the people who have to act on them

What this enables

Academic grounding