AI/ML · Data Engineering · Full Stack
InsightForge
An autonomous data analyst with sandboxed SQL
- Status
- Complete
- Year
- 2026
- Live application
- Runs locally with Ollama. No authentication; not publicly deployed.

- unsafe-SQL rejection and injection safety
- 100%
- unsafe-SQL rejection and injection safety
- numerical correctness with a local 4B model
- 66.7%
- numerical correctness with a local 4B model
- layers of SQL isolation
- 4
- layers of SQL isolation
01 Problem
"Chat with your data" tools risk unsafe execution, wrong analysis and prose containing numbers nobody computed. InsightForge treats the model as a planner whose proposals are validated, runs everything in an OS-level sandbox and checks every number in the explanation against computed facts.
02 Product overview
A LangGraph agent proposes typed statistical operations first and SQL second, with up to three repair attempts. Model-written SQL passes four layers of isolation before it touches data. Explanations are written with fact placeholders that a verifier fills from computed results, and charts are built server-side with Plotly.
Implemented
- Four-layer SQL isolation: SQLGlot AST policy → isolated Python worker → sandbox-exec / bubblewrap → locked DuckDB
- 12 typed statistical operations
- Fact-placeholder explanations checked by a number verifier
- Ingestion of CSV, TSV, XLSX, Parquet and SQLite into versioned Parquet with profiling
- Chart Studio, Monaco query console with an audit log, saved analyses and export
Planned, not yet built
- Warm sandbox pool
- PostgreSQL connector
- PDF export
- Larger-model evaluation
- Cross-filtering
03 System architecture
04 Core technologies
Analysis
- DuckDB
- SQLGlot
- pandas
- NumPy
- SciPy
- statsmodels
- scikit-learn
Agent
- LangGraph
- Ollama (qwen3:4b-instruct)
Application
- FastAPI
- Pydantic 2
- SQLAlchemy 2
- Next.js 16
- Plotly
- Monaco
05 Engineering challenges
Streaming through Next.js
Next.js gzip buffered the SSE stream, so progress arrived all at once. A no-transform middleware fixed it.
Isolation that holds
A query filter alone isn't a sandbox. Process-per-query isolation adds roughly 100-300 ms per statement, a deliberate trade for safety.
Numbers you can trust
Explanations reference facts by placeholder, so the model cannot invent a figure that wasn't computed.
06 Technical decisions
- 01Typed operations first, SQL second
- 02Process-level sandbox rather than query filtering
- 03Parquet as the canonical store
- 04Server-built charts
07 Verified test results
- pytest
- 170 passed, 1 skipped
- Playwright E2E
- 6/6
- Vitest
- 12/12
Latest GitHub Actions run on main: passing. Counts as reported in the repository's README and testing docs.
08 Screenshots
09 Links
No public deployment. Runs locally with Ollama. No authentication; not publicly deployed.


