PrivaLens DataRescue
Inspiration
Data is everywhere, but real-world datasets are rarely clean, consistent, or safe to use. Missing values, duplicates, inconsistent formats, and privacy risks can quietly turn a valuable dataset into a liability.
We were inspired by the idea of treating data preparation not as a manual cleaning task, but as an autonomous rescue mission.
Instead of asking users to find problems one by one, we built PrivaLens DataRescue to continuously inspect data, understand what is wrong, decide what can safely be fixed, and verify that the fixes actually worked.
Our goal was simple:
Turn messy and risky data into trustworthy, ML-ready data — autonomously.
What it does
PrivaLens DataRescue is an autonomous AI data guardian that analyzes datasets for data-quality, privacy, and machine-learning readiness issues.
It can detect:
- Missing values
- Duplicate records
- Inconsistent formats
- Malformed dates
- Inconsistent categories
- Invalid values
- Unit inconsistencies
- Outliers
- Direct identifiers
- Quasi-identifiers
- Re-identification risks
- Privacy vulnerabilities
- ML-readiness problems
Instead of blindly modifying the dataset, DataRescue creates a rescue plan and classifies proposed actions according to their risk.
Safe operations can be performed automatically, while privacy-sensitive or potentially destructive operations can require human approval.
The most important part is the verification loop:
Detect → Reason → Plan → Act → Verify → Adapt
After making changes, the agent re-runs quality and privacy checks, evaluates utility loss, and can reconsider its strategy if the result is not good enough.
The result is a cleaner, safer, and more trustworthy dataset accompanied by an auditable record of what happened.
How we built it
We designed PrivaLens DataRescue as an agentic orchestration layer around deterministic data and privacy analysis engines.
The architecture combines:
- AI agent orchestration for planning, prioritization, reasoning, and tool selection
- Deterministic data-quality analysis for reliable metrics
- Privacy analysis and attack simulation to measure actual risk
- Automated mitigation tools for safe transformations
- Human-in-the-loop approval for higher-risk operations
- Validation and re-testing after every significant transformation
- Persistent job state for long-running rescue workflows
- Dataset versioning to preserve the original data and enable rollback
- Audit logging for traceability and accountability
- Interactive dashboards for monitoring the agent's progress and results
We deliberately separated AI reasoning from numerical computation. The agent can decide what should be done, but important metrics such as privacy scores, attack results, and utility measurements are calculated by deterministic systems rather than invented by the LLM.
The system follows:
Dataset
↓
Profile
↓
Detect Problems
↓
Prioritize
↓
Generate Rescue Plan
↓
Execute Safe Fixes
↓
Request Human Approval When Needed
↓
Apply Mitigations
↓
Attack & Re-test
↓
Measure Utility
↓
Verify
↓
Rescued Dataset
Challenges we ran into
One of the biggest challenges was making the project genuinely agentic rather than simply connecting an LLM to a data-cleaning script.
A real autonomous system needs to understand the current state, choose appropriate actions, observe their results, and change course when necessary.
Another challenge was balancing privacy and utility. Making data more private can reduce its usefulness for analysis or machine learning. We therefore designed the system to look for the least-destructive mitigation that achieves the required privacy improvement.
We also had to think carefully about trust. An autonomous system modifying datasets can be dangerous if its decisions cannot be explained or reversed. This led us to introduce human approval gates, policy-based decisions, dataset versioning, audit trails, and post-action verification.
Finally, we had to make the experience reliable enough for a live demonstration while still representing a realistic production architecture.
Accomplishments that we're proud of
We are particularly proud of turning a traditional data-cleaning workflow into an autonomous rescue loop.
The system does not simply tell the user that something is wrong. It can:
- Discover the problem.
- Explain why it matters.
- Prioritize it.
- Propose a solution.
- Decide whether it is safe to automate.
- Execute the transformation.
- Re-test the resulting dataset.
- Measure the privacy and utility impact.
- Adapt when the result is insufficient.
- Produce an auditable final result.
We are also proud of the privacy attack → mitigation → re-attack workflow. Instead of claiming that a dataset is private after transformation, the system attempts to demonstrate that the risk has actually decreased.
Most importantly, we built the project around a simple product philosophy:
Don't just analyze data. Rescue it.
What we learned
We learned that building useful AI agents requires much more than connecting an LLM to tools.
The difficult part is designing the control system around the model:
- What should the agent be allowed to do?
- When should it ask a human?
- How should it handle uncertainty?
- How can its actions be verified?
- What happens when a transformation makes things worse?
- How do we preserve user control?
- How do we prevent untrusted dataset content from becoming instructions to the agent?
We also learned that deterministic systems and AI agents work best together. The model provides flexibility and reasoning, while deterministic tools provide reliable measurements and enforce important safety boundaries.
Most importantly, we learned that a convincing agent should not merely produce an answer — it should take action, observe the outcome, and prove that the action worked.
What's next for PrivaLens DataRescue
Our long-term vision is to turn DataRescue into a continuous data guardian rather than a one-time cleaning tool.
Future versions could:
- Monitor datasets continuously
- Detect schema and data drift
- Automatically rescue new dataset versions
- Learn organization-specific data policies
- Support multi-dataset reconciliation
- Build complete data lineage
- Optimize privacy versus analytical utility
- Integrate with cloud storage and data warehouses
- Support scheduled background rescue jobs
- Provide organization-wide governance and observability
Ultimately, we want PrivaLens DataRescue to become the layer that sits between raw data and trustworthy AI, ensuring that data is not only clean enough to use, but also safe, explainable, verifiable, and ready for machine learning.
From messy data to trusted intelligence.
Log in or sign up for Devpost to join the conversation.