01
Black-box predictions
A prediction can rank risk without clarifying which intervention might change the outcome.
Open-source private-use prototype
Causal AI prototype for private analysis and learning
aitiolin is a browser-based interface for interactive DAG building, identification analysis, causal effect estimation, robustness testing, and time-series causal analysis. It helps users move from “what correlates with churn?” to “what may happen if we actually change something?” in a transparent, non-production setting.
- Interactive graph canvas with human-reviewed assumptions
- Identification checks before estimating an effect
- Robustness dashboard for stress-testing results
- White-box behavior for explainable analysis
Identification panel
Backdoor paths, admissibility checks, and adjustment sets.
Illustrative query: Online Security → Churn
Identification is conditional on the proposed DAG. This panel is an illustration, not a recorded analysis.
Identification is conditional on the proposed DAG. This panel is an illustration, not a recorded analysis.
Robustness
Parallel estimators, placebo tests, subsamples, and sensitivity sweep.
Evidence first: compare estimators, inspect refutations, and review sensitivity.
A composite score is optional and is not a probability of causal correctness.
A composite score is optional and is not a probability of causal correctness.
Why this matters
Causal analysis should make assumptions visible before interpretation
Many AI projects produce scores without clearly showing why a result appears or whether the underlying assumptions are defensible. For private users and researchers, that makes intervention reasoning harder, not easier.
02
Correlation is not intervention
Any intervention strategy needs a causal question: what happens if we change pricing, support, or retention spend?
03
Shared assumptions matter
Visual graphs allow domain experts, analysts, and other reviewers to examine assumptions together.
Prototype
One open-source GUI prototype for the full causal workflow
Instead of stitching together multiple libraries and notebooks, this open-source prototype concentrates graph drafting, identification, estimation, and robustness into a single browser-based workflow for private use.
A
Interactive DAG builder
Drag variables onto a canvas, create directed edges, and persist graph state across sessions.
B
Graph Copilot
Get AI-assisted edge suggestions with reviewable evidence and accept/reject decisions.
C
Identification analysis
Check DAG validity, admissibility, adjustment sets, and candidate identification strategies.
D
Effect estimation
Run estimators such as linear regression, propensity score matching, and weighting.
E
Robustness dashboard
Compare estimators, inspect refutations and sensitivity, and review alternative causal models. The summary score is secondary.
F
Time-series mode
Analyze rolling-window stability where a time dimension is available.
Workflow
From CSV to causal insight in eight steps
The workflow is structured so users inspect assumptions before effect estimation and robustness before interpretation.
Load a sample or CSV
Start with a bundled example, or upload a CSV to your running Django server.
Build the graph
Place nodes on the canvas and draw directed edges.
Review model suggestions
Use local heuristics, or preview and approve each external LLM request. Review every proposed edge.
Run identification checks
Validate whether the treatment–outcome question is answerable.
Select treatment and outcome
Specify the intervention question you want to analyze.
Estimate the effect
Run one or more estimators against the causal query.
Stress-test the result
Inspect placebo tests, dummy outcomes, subsamples, and sensitivity.
Export and interpret
Save a run bundle with assumptions, versions and evidence. Use it for learning, not operational decisions.
Try a sample locally
This page is an introduction, not a hosted analysis service. Start the updated Django backend and Vue frontend from the repository instructions. In the application, choose Try a sample and load Telco churn or World Bank health spending. No account or LLM key is required for sample loading.
Review the preset graph and limitations, estimate the effect, then inspect diagnostics and alternative models. Use Reproducibility exports to save the latest recorded stage as JSON or ZIP. Dataset files and row previews are excluded; column names and analysis metadata can still be sensitive.
CSV processing is server-side and is local only when the server runs on your own computer. The privacy panel explains storage, reviewed external requests, retention and Delete session data. Closing a tab alone does not delete server files.
Examples
Two worked examples
Compact demonstrations of the prototype workflow, from a causal question to estimation and robustness review. They are provided for learning and are not business, policy, or medical recommendations.
01
Telco churn
Examines whether longer contracts may reduce churn after accounting for customer differences such as tenure. The walkthrough covers the DAG, adjustment, estimation, refutation tests, and sensitivity.
02
Health spending and child mortality
Explores whether higher health spending relates to lower child mortality after accounting for GDP and region, and shows how functional form and panel data can change interpretation.
Trust and governance
Designed for transparent causal exploration and hypothesis development
aitiolin is developed as an open-source prototype for private experimentation and learning, not for autonomous or operational decision-making. The graph, identification logic, and robustness checks are visible and reviewable.
aitiolin can support transparent, human-reviewed workflows, but legal compliance depends on the use context. This tool is intended for private use only
and is not intended for business decisions.
W
White-box behavior
The causal graph makes structural assumptions visible instead of hiding them in a black-box score.
H
Human-in-the-loop
Edge suggestions can be reviewed, accepted, locked, or rejected by a human user.
R
Robustness before interpretation
The dashboard encourages stress-testing before drawing conclusions from an estimate.
News
Latest news
Repository updates for private experimentation and learning. Source version: 0.1.0-prototype.1.
Guided samples, recorded runs and clearer privacy controls
- Load the bundled Churn or World Bank dataset as a sample, with a reviewable hypothesis graph and a short guide. Sample loading needs no account or LLM key.
- Save recorded DAGs, cleaning history, identification results, estimator configuration, seeds, versions and available evidence as JSON or ZIP for reproducibility. Raw CSV rows are not included.
- Keep the data flow explicit: external AI is off by default; each outgoing payload needs review and approval. Inspect server-side storage, session expiry and deletion controls.
- Put evidence first: inspect refutations, estimator comparisons, sensitivity curves and matching alternative-DAG results before the optional composite diagnostic score.
- Run development checks: backend and frontend CI, version checks and a test-gated draft-prerelease workflow are now defined in the repository.
This is an experimental source update, not a production release, security certification or validation of causal conclusions.
Causality Agent introduced
The staged assistant added data profiling, reviewable cleaning, model suggestions, and competing-DAG comparison. In the current version, external LLM requests require explicit enablement and payload approval.
Screenshots
Open-source prototype
Explore the repository and test it privately
aitiolin is an open-source prototype for transparent causal exploration and hypothesis development. It is intended for private use, experimentation, and learning, not for production or business decision workflows.