# 🎯 Sales Trainer

A **corporate, multi-user sales-training simulator**. Admins upload/describe a product; the app
analyzes it and generates **15 realistic customer personas** (5 per buying-intent tier) with real,
varied pains. Trainees **chat one-on-one** with each persona to practice closing a sale — customers
negotiate, stall, and refuse unless their pain is genuinely resolved. Debrief reveals + coaches.

Built on patterns from the **CrowdSight / MiroFish** swarm engine and clean-room Hermes Brain & Tools plugin.

---

## Features

- **Login + roles** (no self-registration): `super_admin` / `admin` / `user`.
  - Admin builds & edits **persona groups** (product + 15 personas), hand-edits personas, sees analytics.
  - User (trainee) **can't create** — only selects a group and practices; sees own results.
- **Input** via form (product / segment / description) **and/or file upload** (.pdf/.md/.txt).
  Product data is used mainly to extract **pains**; personas are reusable across similar products.
- **15 personas** (5 × tier A/B/C):
  - A = ready to buy · B = unsure · C = not interested but has pain (hardest).
  - Varied demographics, income, occupation, lifestyle, personality — consistent with product.
  - **Pain variety** (directly-solvable / partial / unrelated red-herring).
  - **Negotiation levers** (price, freebies, delivery time, scope, payment).
  - **Initiation mode**: customer opens OR seller must open the sale (outbound, e.g. insurance).
  - **Channel**: Facebook / LINE.
  - Special tier-C **"wrong_text"** persona (appears to buy, loses interest, but still has pain).
- **One-shot rule**: a persona is chatted **once per user** (final); shared across other users.
- **Chat realism**: all tiers can lose; everything negotiates; hidden internal signals + latent
  fields (pain/income/personality/budget) revealed only after the result.
- **Debrief**: short summary + **coaching** (how to answer better on weak-score messages),
  scored by a **separate judge LLM** (no speed factor).
- **Training loop**: win/lose board, **weak-area analysis**, and **user-generated personas**
  (weak-area "lock" or manual form).
- **Admin analytics**: close rate, avg score, hardest personas.
- **EN + TH** UI.

---

## Quick start

### Local (dev)

```bash
# backend (Python 3.11)
cd backend
uv venv --python 3.11 .venv
uv pip install -r requirements.txt --python .venv/bin/python
cp .env.example .env   # edit LLM keys + JWT_SECRET
uv run python run.py   # Flask on :5001

# frontend (separate terminal)
cd frontend
npm install
npm run dev            # Vite on :3000 -> proxies /api to :5001
```

The first run creates a default super-admin: **`admin@salestrainer.local` / `admin123`** (change it!).

### Docker / EasyPanel

```bash
cp .env.example .env   # set LLM_API_KEY + a strong JWT_SECRET
docker compose up -d   # single container serving frontend + API on :5001
```

---

## LLM config

Any OpenAI-compatible endpoint (OpenAI, DeepSeek, or custom base URL):

```env
LLM_PROVIDER=deepseek        # deepseek | openai | custom
LLM_BASE_URL=                # optional override
LLM_MODEL_NAME=deepseek-chat # optional override
LLM_API_KEY=sk-...
```

---

## Architecture

```
frontend/   Vue 3 + Vite SPA (login, dashboard, group builder, personas, chat, debrief,
            gen-persona, weak-areas, analytics). Built to dist/ and served by Flask.
backend/    Flask API (JWT auth, roles, groups, analyzer, persona generator, chat simulator,
            judge, trainee loop, analytics). Filesystem JSON persistence (no external DB).
  app/services/  analyzer · persona_generator · simulator (+ judge) · report · trainee · groups · sessions
docs/PLAN.md   full design & decisions record
```

- **Storage**: `backend/data/` — JSON files per entity (users, orgs, groups, sessions, my_personas).
- **LLM calls**: analyzer (sales kit + pain-fit), persona generator, persona chat, judge.

---

## Tests

Run with the built-in deterministic **mock LLM** (no external key needed):

```bash
cd backend
uv run python scripts/test_m0.py      # auth/roles/no-self-registration
uv run python scripts/test_m1.py      # group create + failure handling + role visibility
uv run python scripts/test_routes.py  # all API routes registered
uv run python scripts/test_e2e.py     # full flow: analyze→personas→chat→debrief→one-shot→board→analytics
```

Real-model verification requires a live `LLM_API_KEY` in `.env`.

---

## Default accounts

| Role | Email | Password |
|------|-------|----------|
| super_admin | `admin@salestrainer.local` | `admin123` (change after first login) |

Admins create additional users (users/login has no self-registration).