# Charie-AI-Data-Analyst

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Sourabh Kumrawat](https://cerebralvalley.ai/u/50ur48h)
- **GitHub:** https://github.com/50ur48h/Charie
- **Demo video:** https://www.youtube.com/watch?v=1fEgyAacReA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2203

Charie is an AI data and business analyst. Anyone can ask a question about their business in plain English, such as "How did we do in December?" or "Which restaurants are losing sales, and why?", and get an answer they can trust and check.

THE PROBLEM
Business owners have the data but not the analyst. Dashboards show charts, not answers, and chatbots that "talk to your data" make up numbers that are hard to verify.

HOW IT WORKS
Charie investigates a question the way an analyst would. It finds the relevant tables, plans what to check, writes SQL, reads the results, and decides what to check next, in several steps when needed. A critic then reviews the draft before Charie answers with key figures, charts, and the exact SQL behind every number. It also says what it could not establish.

HOW META MODELS ARE USED
Every model Charie calls at runtime is a Meta model, through the Meta Model API. Each question takes several Meta calls, each with one job: understanding the question, planning, writing SQL, reading results, reviewing the evidence, and writing the answer. Muse Spark 1.3 runs every step by default, and users can switch any question to Muse Wonder 1.0 in the app. Each Meta model call appears live in the app's Thinking panel with the model name, its job, time and tokens, so the Meta calls are visible as they happen.

SAFE AND HONEST BY DESIGN
- Read-only: Charie proves a database login cannot write before using it, and every query passes a Data Access Layer (a single SELECT, known tables only, limits and timeouts).
- Private: sensitive columns are masked before results reach the model.
- Honest: the critic flags numbers it calculated rather than read and data it did not use, and partial answers are labelled as partial.
- Inspectable: "How I worked this out" shows the method, the findings and the SQL behind each number.

TRY IT (ONLY A META API KEY NEEDED)
- Live demo: https://charie-production.up.railway.app
- Docker: docker run -p 3000:3000 -e META_API_KEY=YOUR_KEY ghcr.io/50ur48h/charie:latest
- Source and step-by-step guide: https://github.com/50ur48h/Charie

The demo comes connected to the open Food Delivery sample database (databasesample.com, 2024–2025, 400,000 orders, 1,000,000 order items), hosted on Neon. Users can connect their own PostgreSQL or SQL Server database in Settings.

BUILT WITH
Meta Model API (Muse Spark 1.3, Muse Wonder 1.0), Python and FastAPI, Next.js and React, PostgreSQL with row-level security, Docker, deployed on Railway.

## More from Meta Global AI Developer Hackathon

- [CoreAudit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2200)
- [Owly](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2201)
- [JunSuzukiJapan](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2202)
- [SENSI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2204)
- [Material Forge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2205)
- [Stochian IT Service](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2206)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2203. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt