Backend & AI Engineer
AI Data Insight Analyzer
Context
Independent technical showcase.
I built this public project to demonstrate how I design AI-backed product workflows, not to present it as paid client work. The goal was to turn a spreadsheet into useful KPIs, charts, findings, and recommendations while keeping model cost, failure modes, and secrets under control.
Engineering approach
The public endpoint is designed as an untrusted boundary.
Key engineering decisions:
- Parse CSV, XLS, and XLSX files server-side and calculate deterministic column statistics before any model call.
- Keep a local analysis path so the product still returns KPIs and charts when AI is skipped, unavailable, or returns unusable output.
- Normalize model-generated KPIs, chart specifications, findings, and recommendations against the real dataset before rendering them.
- Bound public cost with a 1 MB upload limit, 2,000-row and 60-column dataset limits, smaller AI limits, prompt caps, output-token caps, and timeouts.
- Protect the upload path with a signed file-hash challenge and rate limiting.
- Keep provider credentials server-side and expose execution mode, attempted models, token usage, and applied safeguards in the response.
Outcome
Delivered a live, recruiter-testable analyzer that remains useful without a successful model call, keeps public inference bounded, and makes its AI execution path observable rather than hiding it behind a loading spinner.