AI Natural Language Analytics Assistant

Key Result

Plain-English → chart in under 3 seconds

The Challenge

What needed solving

Even with a dashboard (Project 1), store managers couldn't get answers to ad-hoc questions without developer help. 'Which Zara products had the highest markdown loss in Fall?' requires writing a pandas query — something non-technical managers can't do. Additionally, no automated system existed to proactively surface anomalies or compile weekly KPIs, meaning issues like rising return rates went unnoticed for weeks.

The Solution

How we solved it

I built a conversational AI assistant powered by Llama 3.3 70B (Groq). The NL query engine converts plain-English questions into pandas code via structured JSON output, then executes the code inside a RestrictedPython sandbox — only pandas and numpy are accessible; file system, network, and eval are fully blocked. A 10-rule anomaly detection engine runs automatically and surfaces HIGH/MEDIUM/LOW severity alerts live in the chat sidebar. A KPI generator computes ~20 metrics and produces formatted weekly Markdown reports for store managers.

<3s

Plain-English question → Plotly chart

10

Automated anomaly rules (returns, ratings, inventory, revenue)

9

HIGH-severity anomalies detected automatically on first run

100%

LLM-generated code sandboxed — zero filesystem/network access