Systems that prove things about sensitive data, without exposing it.

Where the method comes from

Courts need to know whether a video is real, and whether an algorithm helps the expert who has to testify about it.

• Three UvA degrees: BSc (Hons) in Politics, Psychology, Law and Economics (2016 to 2019), MSc Data Science (2023, thesis on emotion classification from EEG time series), MSc Forensic Science (2025).
• At the Netherlands Forensic Institute I fine-tuned DINOv2 for explainable deepfake detection and tested it against certified examiners. The thesis was defended at the Ministry of Justice and Security.
• ETHDam 2024, my first hackathon: a data bounty marketplace with Waku messaging. Waku prize, $1,000.

• PyTorch
• DINOv2
• Bayesian statistics
• Experimental design
• Chain of custody
• GDPR
• EEG time series
• Solidity
• Waku

For an architectEvidence pipelines that hold up under cross-examination: provenance, reproducibility and explainability are requirements.

For a data engineerCompression-artifact features, evaluation against human baselines, stated error bounds.

For a software engineerTraining code other examiners could run, and a working bounty dApp built in a weekend.

The first sketch of think2earn

Neurotech hardware is closed and expensive, and the people who produce brain data do not own it.

• An EEG data-collection frontend and a plan for manufacturing open-hardware brain-computer interfaces and paying the people who record with them.
• Best Startup and Best Idea prizes, $3,000.
• This is where think2earn started as a hardware project. Every later district builds on it.

• Product architecture
• Incentive design
• EEG
• Frontend prototyping
• Business modelling

For an architectTurning an EEG thesis into a system boundary: device, data, incentives, ownership.

For a data engineerThe collection protocol and the labels were designed before any data existed.

For a software engineerA frontend shipped in days to see whether people would sit through a session.

Prove the inference, hide the brain data

A model says something about your brain. Anyone should be able to check the result without seeing the signal.

• An end-to-end zero-knowledge ML pipeline: fNIRSNet trained with k-fold cross-validation, exported to ONNX, compiled into an EZKL circuit, proof generated and verified.
• ZuThailand fellow on a full-month scholarship. EZKL prize and a main-track honorable mention.
• MegaZU, the EigenLayer and MegaETH builder residency: on-chain attestations for fNIRS data, and a presentation to the judges in a panel with Vitalik Buterin.

$ python zuthailand_ezkl_pipeline.py
Calibration completed.
Compiling circuit... Circuit compilation result: True
Obtaining SRS... SRS obtained: True
Running setup (proving + verifying keys)... Setup result: True
Generating proof... Proof file created at: /content/test.pf
Verifying proof... Proof verification successful: verified

• EZKL
• ONNX
• PyTorch
• k-fold CV
• Halo2 proving
• Attestations
• PySyft
• TenSEAL
• Research writing

For an architectAttest the inference, never publish the data. The later reward systems rely on this pattern.

For a data engineerExport discipline: quantisation and calibration so a network survives becoming a circuit.

For a software engineerA pipeline with an assert after every stage. If it says verified, it verified.

Train on my data without seeing it

Biometric data is sensitive and risky to share. Homomorphic encryption lets a buyer compute on it while the owner keeps the key.

• A bounty marketplace: buyers post data needs and CKKS parameters on-chain; owners encrypt locally with TenSEAL and approve each code request on their PySyft datasite.
• Training runs on ciphertext. Only the owner decrypts the result, and the contract pays.
• Deployed to Polygon and Linea Sepolia testnets with a subgraph, winning the Linea prize and an invitation to the Linea Community Call.

• TenSEAL (CKKS)
• PySyft
• Threat modelling
• Solidity
• The Graph
• Next.js
• Docker
• Linea
• Dynamic

For an architectTrust boundaries drawn so the buyer never holds plaintext, even if the buyer is hostile.

For a data engineerEncrypted feature pipelines, and a working sense of what CKKS can compute and what it costs.

For a software engineerContracts, subgraph, web app and a Python service that ran end to end in 36 hours.

Manufacture the instrument

Universities publish good open hardware and stop at the paper.

• Ocean Protocol research grant, $8,000: I surveyed open fNIRS hardware, wrote the paper and built NIRDuino from Vanderbilt's schematics with the Bowden Lab.
• I ran the hardware QA myself, found hardware bugs and reported them back to Vanderbilt.
• Collected the first public paired light-and-glucose dataset: paired near-infrared readings with continuous-glucose-monitor labels, with each recording signed by the physical unit.

• Hardware QA
• fNIRS optics
• Data acquisition
• Dataset design
• Device attestation
• Technical writing

For an architectThe whole stack from photons to payouts, and where the hardware is the weak link.

For a data engineerA labelled dataset designed before the model: sampling rate, baselines, paired ground truth, provenance per recording.

For a software engineerCapture and attestation tools that run next to real hardware.

From raw light to a metabolic state

Can a cheap open fNIRS device tell fasting from fed? For the person it was trained on, yes.

• Built an end-to-end signal pipeline: turned NIRDuino raw intensities into HbO/HbR, windowed 60-second epochs, extracted statistical and FFT features, and evaluated a logistic regression under stratified 5-fold cross-validation.
• Within-subject accuracy 0.95, 0.96 and 1.00. Across subjects: 0.31 and 0.36. I report the second number because a payout system has to know it.
• Won the Logos prize for NFT-gated decryption on Codex storage, plus a separate prize for the metabolic classifier.

0.95Carol · 5-fold CV

0.96Ron · 5-fold CV

1.00Rob · 5-fold CV

0.31cross-subject transfer

• scipy
• scikit-learn
• Signal processing
• Cross-validation
• NFT access control
• Codex
• Polygon Amoy

For an architectAccess control as a property of the data, not of a server.

For a data engineerThe full signal pipeline, cross-validation that respects the epoch structure, and a negative result stated plainly.

For a software engineerReproducible notebooks with saved artifacts, plus a working encryption gate.

Reward data by its marginal value

Paying for data attracts junk. Pay only for data that measurably improves the model.

• The Vault Berlin, on EigenLayer's Hourglass AVS template: glucose-monitor and NIRDuino data, with on-chain proof that new data improved the model (blocked cross-validation).
• NEAR Developer Rewards, Diamond tier, $17,500: Rust contracts where users stake and earn when their data provably improves performance.
• A Data-Shapley scorer (FastAPI) prices each contribution by its marginal improvement; the next step moves it into a TEE for trustless scoring.

• Rust
• NEAR contracts
• EigenLayer AVS
• Data-Shapley
• Blocked CV
• FastAPI
• TEEs
• Mechanism design

For an architectVerifiable evaluation as the settlement layer: the model's own improvement is the oracle.

For a data engineerAttribution and evaluation protocols that stay honest when money depends on the number.

For a software engineerRust contracts, Python scoring services and AVS operators in one flow.

Prove you were there, without saying where you were

A POAP ties your ENS name to a place and a time, on-chain, forever.

• POAPrivacy mints attendance tokens to single-use stealth addresses (ERC-5564), registered under your ENS name (ERC-6538).
• The keys come from a passkey: WebAuthn signs a fixed message and HKDF derives the spending and viewing keys. No private key is ever shown to the user.
• Finalist (top 2%) and ENS Most Creative Use Case, 2nd place: $10k AWS credits plus $3,000.

Team of four. My part: the passkey-to-stealth key derivation, the address flow and its threat model.

• WebAuthn passkeys
• HKDF
• secp256k1
• ERC-5564 · ERC-6538
• Threat modelling
• Next.js
• Privy
• POAP API
• View-tag scanning

For an architectKey management that removes the key from the user experience: derived on the fly, never stored.

For a data engineerScanning announcements at scale, with view tags as a cheap pre-filter.

For a software engineerCryptographic primitives in TypeScript under a 36-hour clock, and a finalist stage demo.

Proof of sleep, on the phone, in zero knowledge

Sleep trackers collect intimate data to reward you. Prove the outcome instead and keep the readings on the device.

• Eepy Cat reads the gyroscope on-device and scores movement through the night. When you wake up, it generates a Groth16 proof (a Circom LessThan circuit) that the score stayed under the threshold.
• snarkjs runs inside React Native with Node core polyfills. The proof goes through an Express relay and Curvegrid MultiBaas to a verifier on Base mainnet, which mints $EEPY.
• A Next.js simulator models user archetypes against incentive parameters, with a Genkit "AI economist". Curvegrid prize for best use of MultiBaas.

Team build at ETHGlobal Buenos Aires; the repository lives on a teammate's account.

• React Native / Expo
• Circom
• snarkjs · Groth16
• Node polyfills in RN
• Solidity · Hardhat
• MultiBaas
• Base
• Incentive simulation
• Next.js · Genkit

For an architectClient-side proving with a thin relay: the smallest trusted surface between a phone and a chain.

For a data engineerSimulating user segments against incentive parameters before collecting real data.

For a software engineerGetting a ZK prover to run inside React Native.

Open medical hardware needs a different way to fund it

Primavera Hacker invited me to argue that open medical hardware and the way public money is allocated are one problem.

• Talk: "Hacking the State: A Two-Front War for Hardware and Political Sovereignty", 55 slides.
• The argument: new incentive systems for open medical hardware, with think2earn and NIRDuino as the case study.
• Where public health money goes, how to fund hardware without a state budget, and coordination and legislation as engineering problems.

• Public speaking
• Mechanism design
• Open-hardware policy
• Coordination

For an architectSystems thinking applied to institutions: incentives, feedback loops, legislation.

For a data engineerAn argument built from figures about where public money goes.

For a software engineerExplaining hard technical ideas to a room that does not write code.

The systems that run think2earn

Running think2earn solo: AI coding agents draft, and I own the architecture, the review, the tests and what ships.

• Production on AWS: Lambda, Firecracker microVMs, Fargate, S3, IAM, CloudWatch. Linux servers, Docker, Nginx and Caddy, TLS automation, CI/CD, signed Android releases.
• The chat control plane below: one API, three execution lanes, a turn journal that web and mobile both read, a usage ledger settled on every done frame.
• NoCap benchmark: GPT-2 trained on FineWeb with the Muon optimiser, microbatch packing and split-QKV, published with a failure ledger. 2026 grants: Hyperagent Founding 500 ($20k credits) and Thinking Machines Lab ($5k credits).

• AWS Lambda · Fargate · S3
• Firecracker microVMs
• DynamoDB · Aurora Postgres
• Event journal · ledger
• Linux · systemd
• Docker · Nginx / Caddy
• CI/CD · release signing
• GPT-2 on FineWeb
• Muon optimiser
• Benchmark design

For an architectControl planes, isolation and cost attribution as daily work. See the architecture section below.

For a data engineerAn append-only turn journal, a ledger settled transactionally, per-tenant Postgres, and training pipelines that publish their evidence.

For a software engineerInfrastructure, CI/CD and signed releases for real users, maintained solo.

Two leaderboard results on hidden test data

Two competitions had no city. Grants are listed once, in the ledger below.

• 3rd of 12,000+ participants in a repository risk and impact modelling challenge sponsored by the Ethereum Foundation: codebase evolution, maintainer dynamics and dependency risk of core Ethereum repositories, predicting hidden jury data. $3,000.
• EigenLayer Vibescore prediction: predictive models for software-verifiability infrastructure, highest prize on the leaderboard. $3,000.
• Grants (Ocean Protocol, NEAR, Hyperagent, Thinking Machines Lab) belong to projects 05, 07 and 11 and are listed in the ledger.

• Predictive modelling
• Repository analysis
• Feature engineering

For an architectReading a codebase's health from its history: maintainers, dependencies, churn.

For a data engineerTwo leaderboard results on hidden test data.

For a software engineerModels that had to run, score and submit on a deadline.