Agentic AI Decision Engine

What this project does

Agentic AI Decision Engine is a business-oriented demo that turns product and competitor inputs into actionable pricing decisions. It is built for presentations, interviews, and workshops, with a polished UI aimed at executive-style storytelling.

The flow is:

1. Data ingestion: Product from DummyJSON (external catalog simulation) or a manual product (SKU / what-if).
2. Competitor benchmarking: Title + price rows (typed in, or example competitors loaded from the same DummyJSON catalog via API).
3. Decision engine: With OPENAI_API_KEY, you can use LLM reasoning (use_llm: true, or the default when a key is present). Without a key, or with simulated mode, deterministic rules drive pricing, promotion, and inventory guidance using the same JSON shape.
4. Presentation layer: HTML UI at / with a decision hero, confidence, insight, product context, and a visual price comparison; stable contract on POST /v1/decisions.

This is not a replacement for an ERP or enterprise pricing platform. It demonstrates how AI can structure real or simulated inputs into business-ready outputs.

For reviewers

1. Start the API (see Quick start), open http://127.0.0.1:8000/, enter a Product ID, click Fetch product, review competitors (or Load example competitors), choose Use AI (LLM) or Simulated logic, then Get Pricing Recommendation. A guided tour is available in the header.
2. /docs: OpenAPI; try POST /v1/decisions with product_id or product + competitors, and optional use_llm.
3. GET /api/readiness: openai_configured, llm_decisions_ready, rag_ready, etc. (no secrets exposed).

Quick start (local)

Use Python 3.10+ (3.11 recommended; the repo may include .python-version).

python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Optional FAISS index (RAG extension / demos)
python scripts/build_index.py

# Optional: OpenAI embeddings when building the index
export OPENAI_API_KEY=...
export EMBEDDING_MODE=embedding-3-small
python scripts/build_index.py

make dev
# or: uvicorn api.index:app --reload --host 0.0.0.0 --port 8000

Open http://localhost:8000/docs for interactive API documentation.

Example POST /v1/decisions

curl -s http://localhost:8000/v1/decisions \
  -H 'content-type: application/json' \
  -d @- <<'JSON'
{
  "product_id": 1,
  "use_llm": false,
  "competitors": [
    { "title": "Brand A", "price": 100 },
    { "title": "Brand B", "price": 130 }
  ]
}
JSON

The response includes product, competitors, decisions (strategy, recommended price, promotion, inventory, reasoning, and confidence when applicable), source (dummyjson API | manual), and decision_engine (llm | simulated).

Example competitors (DummyJSON)

GET /api/example-competitors/{product_id}?limit=3

Returns suggested peer rows from the same catalog (same category when possible).

Configuration

See .env.example. Important variables:

Variable	Purpose
OPENAI_API_KEY	Enables LLM mode on /v1/decisions (unless use_llm forces otherwise).
OPENAI_CHAT_MODEL	Defaults to gpt-4o-mini.
EMBEDDING_MODE	embedding-3-small or pseudo (offline).
OPENAI_EMBEDDING_MODEL	Defaults to text-embedding-3-small.

Query-time embeddings must stay consistent with how the FAISS index was built.

Deploying (Render)

• Blueprint: render.yaml (install dependencies, run scripts/build_index.py, start uvicorn on $PORT).
• Health: GET /health.
• Set OPENAI_API_KEY as a secret if you use LLM or OpenAI embeddings at build time.

Tests

pytest

Project layout

api/index.py              # FastAPI (demo HTML, decisions, DummyJSON)
app/services/             # ai_decision_engine (LLM), rule_decisions (simulated)
app/integrations/         # dummyjson.py
app/rag/                  # FAISS + retriever (extension)
app/static/demo.html      # Demo UI
docs/images/mvp.png       # MVP screenshot

Notes

• The MVP does not require an external database; the RAG corpus is versioned JSON plus an on-disk index.
• DummyJSON is sample data; production would wire real sources and policies.

---

Creator

Rusben Madrigal
San José, Costa Rica

• LinkedIn: linkedin.com/in/rusmadrigal
• GitHub: @rusmadrigal
• Website: www.rusmadrigal.com