Inspiration
I’ve recently been deeply interested in AI safety, governance, and compliance. As more organizations train models on proprietary or sensitive datasets, a major challenge emerges: companies often need to demonstrate that their datasets comply with safety and regulatory standards, but they cannot reveal the datasets themselves due to privacy, intellectual property, or security concerns.
This led me to explore a simple idea: What if organizations could privately evaluate their datasets using AI, and then generate a cryptographic proof that the compliance decision was computed correctly — without revealing the dataset or the detailed findings?
With the availability of powerful reasoning models like Gemini and modern zero-knowledge proof systems, I saw an opportunity to combine AI-assisted dataset evaluation with verifiable computation, enabling trust without disclosure.
What it does
Noema is a proof-of-concept system for privacy-preserving dataset compliance verification.
A dataset owner uploads a dataset and selects governance policies such as privacy, safety, or regulatory constraints. Gemini evaluates the dataset and produces structured severity assessments for each constraint. These assessments are then used as inputs to a zero-knowledge circuit that verifies that the policy decision (PASS / FAIL) was computed correctly from the declared policy configuration and evaluation outputs.
The resulting proof allows anyone to verify that the compliance result was derived honestly, while the dataset itself — and optionally the individual constraint results — remain private.
How I built it
The system is implemented as a Golang-based backend platform integrating:
- Gemini API for structured dataset reasoning
- A policy evaluation pipeline that converts reasoning outputs into deterministic policy inputs
- gnark for constructing a Groth16 zero-knowledge policy gate circuit
- Cryptographic commitments linking dataset digests, policy configuration, and evaluation outputs
- A lightweight web interface for dataset upload, constraint configuration, evaluation runs, and proof verification
The backend processes datasets, runs AI evaluation, computes policy results, constructs a witness, generates a zero-knowledge proof, verifies the proof server-side, and exposes verification endpoints for external checking.
Prototype note
Noema is an early-stage research prototype demonstrating the architecture for combining AI-assisted dataset evaluation with zero-knowledge policy verification. Core components of the evaluation, policy computation, and proof-generation workflow are implemented, while portions of the UI flow and end-to-end automation are still being refined for demonstration purposes.
Challenges I ran into
Translating policy evaluation logic into circuit constraints required careful design to ensure deterministic witness construction and correct policy evaluation inside the proof system. Working with zero-knowledge circuits in a full evaluation pipeline also required handling constraint modeling, and integration challenges between AI evaluation outputs and cryptographic verification.
In parallel, I experimented extensively with new agent-assisted development workflows, learning how to effectively guide multi-agent coding processes while maintaining reliable and reproducible system behavior.
Accomplishments that I'm proud of
- Building an end-to-end prototype combining AI-assisted dataset evaluation with zero-knowledge verifiable policy computation
- Implementing a working policy gate circuit capable of proving compliance decisions without revealing the underlying dataset
- Successfully bringing a long-standing research-style idea into a functional prototype within a hackathon timeframe
What I learned
This project highlighted how AI reasoning systems and cryptographic proof systems can complement each other: AI provides semantic interpretation of complex datasets, while zero-knowledge proofs provide verifiable guarantees about how policy decisions are computed.
I also gained experience experimenting with emerging AI-assisted development workflows, which significantly accelerated the ability to prototype complex infrastructure concepts.
Overall I had a lot of fun building this project!
What's next for Noema
Future work includes:
- Expanding multimodal dataset support (text, structured data, and images)
- Refining circuit ergonomics and improving proving performance
- Developing stronger standardized governance policy templates
- Exploring verifiable attestation systems for reasoning outputs and evaluation pipelines
Long-term, Noema explores a future where AI dataset governance becomes cryptographically verifiable infrastructure rather than a trust-based process. As advanced multimodal reasoning systems like Gemini gain the ability to analyze massive text, image, and structured datasets for safety, regulatory, and governance risks, the missing layer is verifiability — the ability to prove that compliance decisions were computed correctly without exposing the data itself. Noema demonstrates how AI evaluation combined with zero-knowledge proofs can enable organizations to generate verifiable compliance attestations for their datasets, making large-scale private auditing fast, scalable, and mathematically trustworthy. This paradigm has the potential to become a foundational building block for the next generation of trustworthy AI systems, privacy-preserving data marketplaces, and regulatory-grade dataset governance.
Log in or sign up for Devpost to join the conversation.