Cover
Sasin EMBA · Creative Transformation · Group design project
Sasin School of Management
Group design project · Pitch
SAKSIT logo

Social Analytics & Knowledge-Based Simulation (SAKSIT)

A swarm-agent simulation platform for exploring public response before policy decisions—
demonstrated through the proposed Khlong Toei casino development prototype.

Group 4Team members
  • 2636433Auttapong Budhsombatwarakul
  • 2642837Ekkasit Sitthirotchanawan
  • 2643367Thanapol Kittidulyakan
  • 2641396Bhawat Harnpakdee
  • 2640484Kun Towiwat
  • 2635123Sirinthip Sirikosinaporn
  • 2640850Supanat Hwangkittham
02 · Why the old approach breaks

Some questions are too sensitive to ask—and too important to guess.

Seeking feedback can expose a confidential project. Moving forward without insight can turn a blind spot into a costly failure.

01 · The challenge

A sensitive decision

ASK DIRECTLYRisk exposing a confidential initiative OR DO NOT ASKProceed with costly blind spots
Traditional researchOften captures one point in time—and may not show how reactions develop through social interaction.

When disclosure is risky and uncertainty is costly, decision-makers need a safer way to explore possible reactions.

02 · The SAKSIT / MiroFish solution

AI-assisted social simulation

SAKSIT uses AI to rehearse stakeholder dynamics—not to replace real people or real research.
03 · The benefit

Earlier and safer decision insight

  1. 01Test earlierExplore potential reactions before a sensitive initiative becomes public.
  2. 02See dynamics, not just answersObserve how narratives, influence, resistance, and support may develop over multiple rounds.
  3. 03Focus real researchIdentify the highest-risk assumptions and the most important questions to validate through surveys, interviews, or expert review.

Protect confidentiality, expose blind spots earlier, and direct real-world research toward the decisions that matter most.

85%
External research evidence · not SAKSIT results

Agents matched survey answers 85% as accurately as people matched their own answers two weeks later.

Evidence of potential—not proof that every simulation will achieve the same result.
Where this approach can help
Explore high-impact social and policy decisions.
03 · The Khlong Toei case

Khlong Toei: one proposal, many competing realities.

Decision

Redevelop Bangkok Port land—with or without a casino?

Why it is hard
  • Too sensitive to ask
  • Impact falls unevenly
  • Opinion shifts over time
Expected outcome

Map support, resistance, coalitions and turning points.

Then validate with real people.

Hypotheses · not a poll
Khlong Toei district~104,000 people
DistrictPort landSettlementSchematic · not to scale
1
2,353 raiPort Authority land
2
~40,000 peoplein the settlement · ~232 rai
3
~12,600 householdsto rehouse on ~⅕ of the landJuly 2026 plan
4
Cargo → Laem ChabangPort jobs shift · union ultimatum
Base map simplified from a Wikimedia Commons district map
Public information
  • Livelihoods
  • 33.8%of residents are daily wage-earnersWorkbook · S4 p.5
  • 57.8%of households earn ฿5,000–20,000 a monthWorkbook · S2 p.12
  • Public opinion
  • 70.7%of Bangkokians back moving the portn = 2,500 · affected households not polled
  • 59.2%of Thais oppose both complex optionsNIDA · n = 1,310 · nationwide
  • Information
  • 65%use local TV—the most trusted channelWorkbook · S4 p.7
04 · How the MiroFish workflow operates

From evidence to a living social simulation.

SAKSIT combines human judgment with AI-assisted construction, simulation, and analysis.

Flow · six stepsHuman decidesAI assists / performsEvidence & validation
  1. 1

    Decision question

    Frame the decision, scope and limits.

    Human
    OutputSigned-off brief
  2. 2

    Evidence & seeds

    Collect verified, source-traced evidence.

    HumanAI
    OutputEvidence base
  3. 3

    Knowledge graph

    Map actors, issues and relationships.

    AIHuman
    OutputActor & issue map
  4. 4

    Personas

    Build synthetic composites—not real people.

    AIHuman
    OutputPersona set
  5. 5

    Multi-agent simulation

    Agents interact and shift, round by round.

    AI
    OutputInteraction log
  6. 6

    Report & validation

    Turn results into hypotheses to test.

    HumanAIValidate
    OutputInsight report
OutcomesReport structure · illustrative
6Report delivers
  • Executive summaryMain reactions and open questions
  • Potential risksBacklash and harms that surfaced
  • ConfidenceQualitative signals—never a probability
  • Questions to validateWhat to test with real people
5Simulation reveals
  • Stakeholder positionsWho supports or resists—and why
  • Emerging narrativesWhich stories spread, through whom
  • Turning pointsWhen and why views shift
  • Agreement & conflictWhere coalitions form or split
Human governancePeople own every decisionAI builds and runs the simulation. Humans set the rules, judge the results and stay accountable.
BeforeApprove the inputsDecision question, ethical limits, sources and data rights, persona assumptionsDecision owner · research lead
DuringControl the runSet rounds and triggers, monitor agents, log every intervention, stop implausible behaviourSimulation analyst
AfterValidate before useCheck evidence traces, compare with experts and real data, test critical hypotheses with real peopleResearch lead · expert reviewers
AccountabilityThe final decision stays with named human decision-makers.Simulation output is advisory—never the sole basis for a high-impact decision.
05 · Animated MiroFish demonstration

Watch a synthetic society respond—round by round.

A conceptual demonstration of how evidence becomes interaction, interaction becomes insight, and insight becomes a question for validation.

Concept demonstration — placeholder for real simulation video and reportRound 0 / 6
    06 · Evidence & business value

    Evidence That Grounded AI Personas Can Reflect Real Human Differences

    Stanford HAI, 2025: generative agents built from interviews with 1,052 real people.

    Study design · Park et al., 2024
    0 participantsDiverse by age, gender, race, region, education, ideology
    ≈ 2-hour qualitative interviewLife story and views on contested issues
    Transcript → agent memoryEach transcript became that person's agent context
    People and agents took the same testsSurveys · personality · economic games · experiments
    0%

    Agents matched participants' survey responses 85% as accurately as participants matched their own responses two weeks later.

    GSS · vs. own answers, not raw accuracy · Stanford HAI, 2025
    0%

    Correlation on personality-test results.

    Personality inventory · Stanford HAI, 2025
    0%

    Correlation on behavioral economic games.

    Weakest result, shown as reported · Stanford HAI, 2025
    0 of 5

    Social-science studies successfully replicated by the agents.

    Replayed experiments · Stanford HAI, 2025
    IMPLICATION 01 · Stanford HAI News, 2025

    Rich context outperforms demographic labels

    Interview-grounded agents were more accurate and less biased than agents built from demographics or short self-descriptions.

    IMPLICATION 02 · Stanford HAI Policy Brief, 2025

    Synthetic populations can become decision testbeds

    Explore potential reactions to policies, communications, interventions and events before real-world rollout.

    IMPLICATION 03 · Stanford HAI Policy Brief, 2025

    Validation and governance remain essential

    Agents hold sensitive representations of people: consent, privacy, audit logs, usage controls and the right to withdraw.

    Caution. Stanford demonstrates the potential of well-grounded individual agents. It does not independently validate the predictive accuracy of SAKSIT, MiroFish, or the Khlong Toei simulation. SAKSIT must be calibrated and tested against real Thai data.
    Evidence that grounded AI personas can reflect real human differences
    1,052interviewed participants
    85%of self-consistency on GSS answers
    80%personality correlation
    66%economic-game correlation
    4 of 5studies replicated
    SAKSIT
    SAKSIT tests how a grounded public might respond before you commit.Grounded in Thai data · calibrated before use · governed by design · directional, never a substitute for real consultation.
    Next step: calibrate against three published Thai polls, then pilot one policy and one brand question with a partner.
    07 · Benefits, limitations & human control

    Simulate to learn. Validate to decide.

    AI can reveal possibilities and blind spots, but people remain accountable for the evidence, interpretation, and final decision.

    1Business & decision benefits
    • Earlier insightExplore reactions before public disclosure.
    • Greater confidentialityInvestigate sensitive scenarios with lower exposure.
    • Dynamic understandingObserve interaction and change over time, not only a one-time response.
    • Faster iterationCompare multiple scenarios and interventions.
    • Deeper explanationQuestion synthetic personas to explore the reasons behind a response.
    • Smarter real-world researchFocus surveys, interviews, and expert review on the highest-risk assumptions.

    SAKSIT helps decision-makers surface blind spots earlier, protect sensitive initiatives, and invest real research effort where it matters most.

    2Critical limitations
    • A simulation is not the real population.
    • Results depend on source quality, coverage, and assumptions.
    • Synthetic personas may reproduce bias or stereotypes.
    • Agent behavior may be inconsistent or fabricated.
    • Emergent patterns are hypotheses—not guaranteed forecasts.
    • Simulated support levels must not be treated as polling percentages.
    • Sensitive or personal data must be governed carefully.
    • The system must never be the sole basis for a high-impact public or business decision.
    3Human governance
    Before simulation
    • Define the decision and ethical boundaries
    • Verify sources and data rights
    • Review stakeholder coverage
    • Approve persona assumptions
    During simulation
    • Monitor agent behavior
    • Test alternative assumptions
    • Record interventions and model settings
    • Stop or correct implausible behavior
    After simulation
    • Review evidence traces
    • Compare findings with experts and real-world data
    • Validate critical hypotheses through surveys or interviews
    • Document uncertainty and limitations
    • Keep final accountability with human decision-makers
    Use SAKSIT to rehearse possible futures—not to replace real society.