What if you could ask: “How would different types of Gen‑Z react to this?”
This prototype lets you try that in seconds. Type a message, pick a persona, and get two perspectives:
- Public Reaction — how they’d likely respond out loud
- Internal Thought — what they might actually feel inside
It’s simple, fast, and fun—built from real blog posts, emotion & sentiment models, light clustering, and a tiny retrieval + LLM layer.
- Open the app, choose a persona (e.g., Love, Surprise, Fear, Neutral vibes).
- Paste any statement (news headline, tweet, your scenario).
- Click Find Reaction.
- See the persona’s Public vs Private reactions side‑by‑side...grounded by similar past posts we found.
We don’t send matched text to the LLM...only attributes (emotions/sentiment/masking) once similarity is high—so the reaction stays persona‑aligned without copying anyone’s text.
We keep the UI minimal so anyone can play with the idea quickly.
- Live code (public):
https://github.com/tarashagarwal/genz-persona-simulation/blob/main/src/app/auth/(dashboard)/dashboard/page.tsx
What’s on the page
- A persona picker (4 clusters discovered from data)
- A text box for your input
- A Find Reaction button to call the backend agent
- Two result cards:
- Public Reaction (emotions + sentiment + masking applied)
- Private Reaction (Internal Thought) (emotions only)
- Optional metadata: similarity score, top emotions, and whether masking seems likely
The dashboard is intentionally one page..no maze of routes. Paste text, pick persona, get reactions.
- We start with the Blog Authorship Corpus (Blogger.com): age, job, zodiac, post date, and text.
From date + age, we infer if the author would be Gen‑Z at the time of posting. - We enrich each post with emotions via GoEmotions (RoBERTa) and sentiment via a Reddit‑trained XLNet model.
- If emotion and sentiment disagree, we flag it as masking (a realistic “I look fine, I feel awful” pattern).
- We then discover 4 personas by clustering primarily on emotional tendencies.
- Finally, we build four FAISS indexes...one per persona. When you type something, we:
- search the chosen persona’s index,
- if similarity is high,
- pull its attributes (emotions/sentiment/masking), and
- ask an LLM to write both the Public and Internal reactions in that style.
That’s it...no massive training jobs, just smart reuse of signals already in the data.
- We initially aimed for PANDORA Talks (demographics + personality): https://arxiv.org/pdf/2004.04460
Access required approval; we’ve requested it and may update later. - We used the Blog Authorship Corpus instead: https://u.cs.biu.ac.il/~koppel/BlogCorpus.htm
Helpful fields: post date, age (to infer Gen‑Z), job, and zodiac. - To enrich the text we used:
- GoEmotions (RoBERTa‑base): https://huggingface.co/SamLowe/roberta-base-go_emotions
- XLNet Reddit Sentiment: https://huggingface.co/minh21/XLNet-Reddit-Sentiment-Analysis
This gave us a workable dataset for clustering and persona design.
We didn’t train a new model from scratch (limited time & data). Instead, we:
- Clustered posts into personas using emotion‑forward features,
- Embedded comments and stored them per‑persona in FAISS,
- Matched new inputs by vector similarity (with a cutoff),
- Generated persona‑aligned reactions using an LLM (e.g., gpt‑4.1‑mini) with only the attributes.
Why this route?
- Fine‑tuning wasn’t feasible quickly.
- Vector similarity worked well in early trials.
- It keeps infra light while still feeling persona‑aware.
- Personas largely reflect emotional tendencies (Love, Surprise, Fear, Neutral).
- Zodiac/job/writing style didn’t produce strong clusters here.
- Emotion vs Sentiment can diverge; we treat that as masking.
- LLM tone may soften strong language...true Gen‑Z speech can be spicier.
- Zero‑shot models may miss slang/cultural nuance...fine‑tuning would help.
- Sampling bias: Blogger.com ≠ a perfect Gen‑Z mirror.
Data & Training
- Persona sizes: 3,395 • 17,429 • 58 (sparse) • 75,757
- Needs: bigger, fresher Gen‑Z text; balanced clusters; regular re‑indexing.
Infrastructure
- Cloud GPUs for faster indexing/refresh.
- Automated pipelines for fetch → enrich → cluster → index.
- Cost per request tracking; start serverless, scale to GPU nodes if needed.
UX & Testing
- Define a test set + confidence threshold (≥85%).
- A/B test whether reactions feel relevant, authentic, and engaging.
Summary
Vector search + LLMs can simulate persona reactions today with modest resources.
To ship it as a product, add: larger balanced datasets, Gen‑Z‑aware fine‑tuning, solid retraining pipelines, and ethics & trust checks.
- Rows processed: 96,639
- Discovered personas: 4
- Vibe summary (informal):
- P0 — Love‑forward: affectionate, upbeat, warm
- P1 — Amusement/Gratitude: playful, appreciative
- P2 — Fear/Nervous: anxious, sensitive (very small cluster)
- P3 — Neutral/Info‑sharing: mostly matter‑of‑fact posts
Full technical breakdown (silhouette, feature importances, per‑persona stats) lives in the generated artifacts:
persona_cards.json,persona_assignments.csv,persona_pca_scatter.png.
- Node.js v20.19.1 (LTS) + npm
- yarn v1.22.22
- Python 3.10.9 + pip 25.1.1
- OpenAI API key in
genz-persona-simulation/agentic_logic/.envOPENAI_API_KEY=sk-...
yarn
yarn dev
# http://localhost:3000cd agentic_logic
python build_index.pycd agentic_logic
pip install -r requirements.txt
python -m agent.HRAgent
# http://localhost:5000Any valid email/password format works. Example:
Email: [email protected]
Password: test12345678
| Component | Command | Port |
|---|---|---|
| Next.js Frontend | yarn dev |
3000 |
| Flask HR Agent | python -m agent.HRAgent |
5000 |
- Blog Authorship Corpus — https://u.cs.biu.ac.il/~koppel/BlogCorpus.htm
- GoEmotions blog — https://research.google/blog/goemotions-a-dataset-for-fine-grained-emotion-classification/
- GoEmotions model (HF) — https://huggingface.co/SamLowe/roberta-base-go_emotions
- Reddit Sentiment model (HF) — https://huggingface.co/minh21/XLNet-Reddit-Sentiment-Analysis
- Processed dataset (HF) — https://huggingface.co/datasets/tarashagarwal/genz-persona-simulation
- Frontend page (public) — https://github.com/tarashagarwal/genz-persona-simulation/blob/main/src/app/auth/(dashboard)/dashboard/page.tsx
genz-persona-simulation/
├─ agentic_logic/
│ ├─ agent/
│ │ └─ HRAgent.py # Flask app entry (python -m agent.HRAgent)
│ └─ .env # OPENAI_API_KEY=...
├─ data_processing_code/
│ ├─ BuildBlogsData.py
│ ├─ BuildBlogDataWithSentiments.py
│ └─ get_personas.py
├─ persona_assignments.csv # (generated)
├─ persona_cards.json # (generated)
├─ persona_pca_scatter.png # (generated)
└─ README.md