Insightoverflow-ai

Inspiration

Data teams told us the same story: ingestion is automated, but transforming raw records into narratives still requires SQL wizards and days of iteration. With Google Cloud and Vertex AI raising the bar for generative analytics, we wanted a copilot that could deliver grounded insights the moment fresh data lands in BigQuery.

## What it does

InsightFlow AI ingests events through a Fivetran-style connector, stores them in BigQuery, and exposes a FastAPI service that asks Gemini to generate SQL and executive summaries. Elastic hybrid search supplies contextual snippets, and a Streamlit UI lets business users chat with their metrics, inspect the SQL, explore returned rows, and browse relevant knowledge-base hits.

## How we built it

We implemented a custom connector scaffold that mimics Fivetran stateful syncs, then layered a FastAPI backend with pluggable clients for BigQuery, Elastic, and Vertex AI so each can fall back to offline stubs during local testing. Prompt templates keep Gemini focused on the configured table, and summaries use Elastic results for grounding. A Streamlit frontend orchestrates the calls, while Dockerfiles, docker-compose, and a Cloud Run deployment script make it portable.

## Challenges we ran into

Application Default Credentials on Cloud Run required revisiting our “credentials or bust” guards so services stay online without explicit key files. Gemini initially hallucinated extra tables, forcing stricter prompt fences and parsing. Latency crept above four seconds until we cached metadata and throttled Elastic payloads. Finally, we needed believable data for demos, so we scripted synthetic inserts and a repeatable BigQuery bootstrap.

## Accomplishments that we're proud of

• End-to-end flow from connector sync to executive summary running fully serverless on Cloud Run.
• Resilient offline mode that lets pytest run without cloud dependencies while keeping the real integrations ready for production.
• A UX that surfaces raw rows, SQL, and context side by side, building trust in every response.
• Deployment tooling that gets a fresh environment live in minutes with minimal manual steps.

## What we learned

Balancing hackathon velocity with production hygiene is all about graceful degradation: every external service needs a fallback and clear logging. Prompt engineering is iterative—small wording tweaks dramatically change Gemini’s reliability. And grounding LLM outputs with search context is critical; even simple Elastic filters boost answer quality.

## What's next for Insightoverflow-ai

Next we’ll add automatic SQL validation against table schemas, richer data quality checks before summarization, and scheduled connector runs so BigQuery stays fresh. We also want to publish hosted demos (UI + API) with live datasets, integrate Vertex AI feedback signals to measure answer usefulness, and ship a Terraform blueprint so teams can stand it up in their own projects quickly.

Built With

• bigquery
• docker
• dockercompose
• elasticsearch
• fastapi
• fivetran
• google-cloud
• http
• pydantic
• pytest
• python3.11
• streamlit
• vertex-ai
• vertexai