carol calin

amsterdam · researcher & builder

I work at the intersection of biometric hardware, machine learning, and verifiable on-chain systems. I build open-source tools, win hackathons, and think about what machines can and cannot perceive. Currently a solo founder shipping open neurotech with AI coding agents as my engineering team.

what i'm doing now

• active think2earn Open-source brain-computer interface · solo founder Manufacturing NIRDuino open-hardware near-infrared spectroscopy devices (in collaboration with Vanderbilt University). Building a coordination protocol where users earn rewards for contributing biometric data that demonstrably improves ML models. Shipping with AI coding agents (Claude Code, Gemini, Devin, Cursor).
• summer 2025 NEAR Protocol Rewards Program Diamond tier · $17,500 grant · 28 commits / 22 PRs Building a staking mechanism on NEAR where users lock funds and earn rewards when their contributed data provably improves model performance. Extending The Vault Berlin work.
• ongoing AI Safety & Evaluation Designed a custom emotion-regulation safety eval using the AISI inspect-ai framework. Quantitative measurement of how effectively frontier LLMs help users regulate emotions. Double-blind, controlled, longitudinal trial design with multiple model conditions.

selected work

• 2023–2024 Deepfake Detection Research Netherlands Forensics Institute (NFI), The Hague · Digital & Biometric Traces Division Fine-tuned vision transformers (DINOv2) on deepfake datasets using frequency-domain features. Conducted a comparative study against trained human forensic analysts — model performance showed no statistically significant difference. Manufactured deepfakes using FaceFusion and GPUs for controlled evaluation. Thesis defended at the Dutch Ministry of Justice & Security; recommendations directly informed national forensic and AI-safety strategies.
• 2024 Deep Funding / pond.xyz — 3rd Prize 12,125 participants · $2,000 prize · sponsored by Vitalik Buterin Built ML models evaluated against pairwise comparisons from a human jury to assign contribution weights across 45 core open-source repositories in the Ethereum ecosystem. Published writeup on Ethereum Magicians forum.
• 2021 Community Manager — LevX (ex-Sushi) Independent DeFi tooling developer Moderated Discord, handled user support for smart contract interactions on Etherscan, built a chatbot, authored FAQ documentation, organized community calls.

hackathons & prizes

• ETHGlobal Cannes 2025 Finalist · ENS Most Creative 2nd POAPrivacy — stealth address POAP claiming (ERC-5564/6538) so attendees can prove attendance without linking identity to physical location. Passkey-based key derivation via WebAuthn + HKDF.
• ETHDam 2025 Waku Prize Data Bounty Marketplace — smart contract for posting and claiming bounties on datasets, with decentralized messaging via Waku.
• Zuitzerland Logos Prize · Track Prize NFT-gated decryption for Codex storage — files only decryptable by wallets holding a specific NFT on Amoy Polygon Testnet. Also won separately for NIRDuino metabolic state classifier.
• ZuVillage Georgia Best Startup · Best Project EEG data collection frontend + business plan for manufacturing open hardware BCI devices. Starting point for think2earn as a hardware project.
• ZuThailand EZKL Prize · Honorable Mention First end-to-end ZKML pipeline: EZKL proofs over a BCI model so inference results are verifiable on-chain without revealing raw data.
• The Vault Berlin EigenLayer Modified EigenLayer's Hourglass AVS template. Collected CGM + NIRDuino data, proved on-chain that new data improved model performance. Presented to EigenLayer and MegaETH judges.
• MegaZU Builder Residency Month-long residency hosted by EigenLayer + MegaETH. Presented to MegaETH and EigenLayer judges in a panel alongside Vitalik Buterin. On-chain attestations for fNIRS data, experiments with homomorphic encryption (PySyft, TenSEAL) for training on private biometric data.
• ETHGlobal Buenos Aires Curvegrid Best Use of MultiBaas Eepy Cat — privacy-first proof-of-sleep mobile app. Gyroscope data stays on-device; Groth16 ZK proof of "stillness" submitted via Curvegrid MultiBaas to a Solidity verifier that mints reward tokens.
• ETHGlobal Bangkok Linea 2nd Place Biometric data sale contracts on Linea. Homomorphic encryption (PySyft, TenSEAL) with federated training. Invited to Linea Community Call.
• NEAR Protocol Diamond Tier Grant ($17.5k) Staking mechanism: users lock funds, earn rewards when contributed data provably improves model performance.
• Ocean Protocol Research Grant ($8k) Surveyed open hardware fNIRS devices, manufactured NIRDuino from published schematics, collected first public paired light-and-glucose dataset.
• EigenLayer AI Competition Winner ($3,000) Won EigenLayer's AI competition with think2earn.
• EigenLayer Vibescore Top 5 AI prompting competition with the @think2earnBCI account.

projects

• 2025 Fileverse Multi-Agent Knowledge Base Designed and deployed a decentralized shared memory system for coordinating multiple AI agents. Cloudflare Worker proxy over Fileverse ddocs with SHA-256 concurrency control, semantic translation rules for visual content, and a vision for an Agentic Messenger product.
• 2024–2025 Beety — On-Chain Data Reward Loop Users upload CGM data; app creates an EigenLayer AVS task, checks whether data improved the model (blocked cross-validation), and lets users claim token rewards. RandomForest glucose prediction benchmarked against medical-grade CGM.
• 2024 agent_logic — Autonomous Impact Evaluator FastAPI service with Data-Shapley scorer that prices each data contribution by marginal model improvement. Migrating to run as a NEAR Shade Agent inside a TEE (Phala) for trustless scoring and payouts.

education

• 2023–2025 MSc Forensic Science University of Amsterdam Thesis: Explainable Deepfake Detection Methods in Videos, defended at the Dutch Ministry of Justice & Security. Also classified fraudulent Ethereum smart contracts using opcode data and ML. Elective: Evolutionary Algorithms.
• 2021–2023 MSc Information Studies: Data Science University of Amsterdam Thesis: Emotion classification from EEG brain signals (DEAP dataset). Electives: Modelling System Dynamics & The Social Web.
• 2018–2021 BSc (Hons) PPLE — Psychology University of Amsterdam Politics, Psychology, Law & Economics. Programme committee member. Amsterdam Data Science & AI Minor (30 EC).

proficiencies

writing & thinking

Universities publish world-class open hardware and then stop, because academic incentives end at publication. I finish the job — productize the device, collect the dataset, train the model, ship the app.

The mission: get open-source neurotech to the level of Neuralink — built by agents + humans, not a closed billion-dollar lab. Agent-first is the only way a small team reaches that frontier.

What competitors miss: Dexcom, Abbott, Oura and Whoop can't ship user-owned, edge-inference agents without cannibalizing the cloud-data and subscription revenue that funds them. Agents are how a small, open team out-iterates both — and the open dataset compounds into a moat forks start from zero against.