HyperPersona

Synthetic JTBD interviews and adversarial hypothesis testing for early product discovery.

HyperPersona is an open-source research and engineering project exploring how synthetic behavioral personas can help product teams expand and stress-test hypotheses before recruiting human participants.

The project is aimed at early problem-solution discovery / Jobs-to-be-Done (JTBD) research across B2C digital products, B2B/SaaS, and physical products.

Synthetic evidence is not customer evidence. HyperPersona generates possibilities and challenges hypotheses; human research is required for external validity.

MVP

The current v0.1.0 engineering slice implements:

Research brief
    |
    v
Brief normalizer
    |
    v
Behavioral cohort (12 personas by default)
    |
    v
Independent JTBD interviews
    |
    v
Atomic observations
    |
    v
Candidate hypothesis
    |
    v
Six-agent adversarial panel
    |
    v
Evidence graph

The six adversarial roles are:

• NULL — asks whether the problem is trivial or absent.
• INCUMBENT — asks whether existing solutions already solve the job.
• BEHAVIORAL — searches for behavior inconsistent with the proposed mechanism.
• CONTEXT — identifies circumstances where the problem disappears.
• ALTERNATIVE_CAUSE — proposes competing causal explanations.
• BOUNDARY — narrows where the hypothesis should and should not generalize.

Important limitation

The MVP currently uses deterministic local logic and a MockLLM abstraction. It is intended to validate orchestration, schemas, provenance, persona isolation, interview flow, adversarial routing, and evidence-graph construction before a production LLM is connected.

Do not interpret current generated interviews as empirical consumer research.

Quick start

Requirements: Python 3.11+ recommended.

git clone https://github.com/joinamber/Hyperpersona-research.git
cd Hyperpersona-research

python3 -m venv .venv
source .venv/bin/activate

python -m pip install --upgrade pip
python -m pip install -r requirements.txt
python -m pytest -v

Start the API:

uvicorn app:app --reload

Then open:

http://127.0.0.1:8000/docs

FastAPI provides an interactive Swagger interface for exercising the MVP endpoints.

API workflow

A typical study runs through these endpoints:

POST /v1/studies
POST /v1/studies/{sid}/brief/validate
POST /v1/studies/{sid}/cohort?n=12
POST /v1/studies/{sid}/interviews/run
POST /v1/studies/{sid}/extract
POST /v1/studies/{sid}/adversarial/run
GET  /v1/studies/{sid}/evidence-graph

Example study payload:

{
  "domain": "B2B_SAAS",
  "target_actor": "SMB employees",
  "target_situation": "submitting work expenses",
  "problem_hypothesis": "Employees need help because receipts get lost",
  "discovery_goal": "discover unmet jobs and friction",
  "market_context": "UK SMEs"
}

Research principles

HyperPersona is designed around several constraints:

1. Behavior before demographics. Personas vary primarily on situations, behaviors, constraints, incumbent solutions, experience, and workarounds.
2. Independent exploration. Persona interviews are kept separate to reduce synthetic conformity.
3. Open discovery before convergence. The system is intended to surface unknown hypotheses before structured comparison.
4. Counterexamples are first-class evidence. Some personas should experience little or no meaningful problem.
5. Provenance is mandatory. Observations retain references to persona and interview turns.
6. Synthetic counts are not prevalence estimates. A frequency within a generated cohort must never be presented as a percentage of real customers.
7. Adversarial review is not validation. Internal robustness remains distinct from evidence observed in humans.

Repository structure

Hyperpersona-research/
├── app.py
├── requirements.txt
├── README.md
├── CONTRIBUTING.md
├── LICENSE
├── docs/
│   └── ARCHITECTURE.md
└── tests/
    └── test_pipeline.py

See docs/ARCHITECTURE.md for the current architecture.

Current limitations

• In-memory storage only; restarting the process clears studies.
• Deterministic simulated persona answers rather than production model calls.
• Fixed interview questions rather than dynamic JTBD laddering.
• One candidate hypothesis generated per extraction run.
• No embedding-based persona deduplication or insight clustering yet.
• No human-validation ingestion/calibration layer yet.
• No cross-model stability testing yet.

Roadmap

Near-term research and engineering priorities:

• production LLM gateway with structured outputs;
• persona diversity and semantic deduplication;
• dynamic JTBD interviewer with bounded laddering;
• atomic evidence extraction and clustering;
• persistent study/evidence storage;
• prompt/model versioning and observability;
• bias and demographic-dependence audits;
• human-validation workflow;
• controlled comparison against generic LLM-assisted discovery.

Contributing

Contributions are welcome. See CONTRIBUTING.md.

Changes to persona generation, interviewing, evidence extraction, adversarial reasoning, or scoring should state both the expected technical effect and the research-method implication.

License

HyperPersona is licensed under the Apache License 2.0. Commercial use, modification, and redistribution are permitted subject to the license terms.

See LICENSE.

Project status

HyperPersona is an experimental research prototype. Its purpose is to investigate whether synthetic-persona workflows can improve the breadth, falsifiability, and efficiency of early product-discovery hypotheses—not to replace human user research.