Back to work
Confidentiality note: Screenshots use sample or non-sensitive data. Employer-owned source and private datasets are not published.
Multi-agent analytics
Data Project Analyzer
A self-service analytics platform that ingests business data, builds context, generates dashboards, and answers grounded questions through deterministic analytics plus LLM agents.
2h -> <5m reporting18-agent pipelineGrounded chat + charts
Problem
Internal teams needed faster answers from CSV, Excel, JSON, and operational datasets without waiting for repeated manual reporting cycles.
Approach
- Separated deterministic analytics from LLM reasoning so simple calculations remained fast, repeatable, and inexpensive.
- Used agent orchestration to build dataset context, route questions, produce visualizations, and explain why a metric changed.
- Added streaming feedback and code execution so analysts could ask follow-up questions without leaving the workflow.
Architecture
- Upload layer normalizes CSV/XLSX/JSON inputs.
- Deterministic compute layer handles profiling, KPIs, transformations, and chart data.
- LLM agents build context, answer questions, and select tools.
- Streaming UI returns progress, charts, and explanations in one loop.
Production
- Model routing keeps lightweight tasks away from expensive calls.
- Structured outputs reduce brittle parsing.
- Cached context keeps repeated questions responsive.
Result
The platform turned reporting from a multi-hour workflow into an interactive analysis loop measured in minutes.
Stack