Senior Machine Learning Engineer

Tools for Humanity (TFH) designs and builds technology behind World. World is building a real human network designed to accelerate people in the age of AI. As bots and autonomous agents reshape the internet, people, institutions, and applications need a trusted way to confirm who is a real human while preserving privacy. The TFH and World tech stacks make this possible: the Orb verifies real, unique people, World ID proves it privately, and World App puts these capabilities, and more, in people’s hands. Together, they add a human layer to an AI-driven internet. World is already running at a global scale. More than 17 million people across 160 countries have verified with World ID, and more new Orb verifications take place each week. World App is already among the most used wallets globally. Developers are integrating World ID to build safer online experiences and create spaces where real people can participate, earn, and be recognized in ways AI simply can’t replicate. Founded in 2019, TFH has more than 400 people across hardware, software, AI, cryptography, mobile engineering, and global operations. Our teams come from OpenAI, Tesla, SpaceX, Apple, Google, Stripe, Meta, Coinbase, Palantir and MIT Media Lab. We’re backed by leading investors, including a16z, Khosla Ventures, Bain Capital Crypto, Blockchain Capital, Variant, Tiger Global, and Coinbase Ventures, as well as prominent operators and founders across fintech and AI. TFH and World have been featured on the cover of TIME Magazine, highlighted in Fast Company’s Next 5 in Fintech, and explored in a Bloomberg deep dive. The New York Times, Bankless and TechCrunch have all recognized our collective progress in identity, cryptography, AI, and global-scale hardware deployment. Our leadership is also named to the Time AI 100. Learn more about the newest product launches from our Liftoff event. The AI & Biometrics team at Tools for Humanity owns the machine learning systems behind the World Network. Our iris and face recognition systems, our anti-spoofing pipeline, and the models running on the Orb and on phones are what make Proof of Personhood actually work at scale. Within the Face team, we develop systems for face verification, uniqueness and duplicate detection, presentation attack detection, and supporting capabilities such as face detection, image quality assessment, occlusion detection, pose estimation, and others that allow these systems to operate reliably in real-world conditions. Our work spans the full machine learning lifecycle: defining data-collection and labeling requirements, curating datasets, conducting applied computer-vision research, training and evaluating models, adversarial testing, supporting deployment, investigating production behavior, and monitoring performance after release. The problems we work on are grounded in real-world conditions. Our systems must handle variation in cameras, lighting, pose, image quality, occlusion, user behavior, and attack methods. They must also operate within strict latency and memory budgets and interact with privacy-preserving systems (Anonymized Multi-Party Computation) that compare users against a growing identity set. At our scale, small regressions can have a meaningful impact, so evaluation, operating-point selection, and production monitoring are central to how we work. We are pragmatic about our methods. We use deep learning where it provides clear value, classical computer vision and image processing where they are more efficient or reliable, and hybrid approaches when they produce the best system. The team is well supported. We have a dedicated Mobile subteam that owns deployment to mobile devices, ML Infrastructure and MLOps subteams that own the data pipelines, training clusters, GPU fleet, and much of the underlying tooling, and Face and Iris subteams dedicated to each modality. Face engineers own problem formulation, data requirements, data-quality decisions, model behavior, evaluation methodology, and investigation of model failures for their projects to succeed. We also work closely with the Orb software, Mobile, Proof of Personhood, and Product teams to take ideas to production. In this role, you will: Own face machine learning projects from initial problem definition through data preparation, experimentation, evaluation, production validation, and monitoring, while partnering with different teams. Improve our core biometric identification and anti-spoofing models, training and iterating on deep learning architectures, losses, and data pipelines, with model size, latency, and memory budgets as first-class design constraints from day one. Use classical computer vision and image processing when they are the right solution, whether as a complete method, a preprocessing stage, a lightweight on-device component, or a diagnostic tool. We are not a deep learning only team, and on-device compute rewards the discipline of using the lighter solution when it fits. Lead independent applied ML initiatives end-to-end: form a written hypothesis, design ablations that isolate variables, run experiments, read results honestly, and know when to ship and when to stop chasing the last percentage point. Work directly with face images and datasets: improve collection and labels, inspect difficult samples, identify failure modes, and use those findings to drive model and evaluation changes. Build evaluation pipelines that catch model regressions before they reach production, monitoring that can reveal data drift, score-distribution changes, new attack patterns, cohort-specific regressions, or unexpected model behavior, and improve internal tools for data analysis, model training, evaluation, visualization, red teaming, and monitoring. Read and evaluate relevant computer vision and biometrics research, adapting useful ideas to our systems without treating paper reproduction as the objective. Write design documents, experiment reports, post-launch analysis and technical proposals that remain useful after the project is complete, and others can understand and build on your work. Help shape technical standards across the AI & Biometrics team: evaluation methodology, experimentation discipline, model versioning, monitoring, and across any engineering and scientific standards of the team.

Qualifications

Significant hands-on experience training, evaluating, and shipping deep learning systems for computer vision, with practical understanding of latency and memory constraints, Experience taking an ambiguous machine learning problem and turning it into a structured technical plan, Strong practical knowledge of model training, including data pipelines, augmentations, architecture selection, loss functions, optimization, hyperparameter tuning, and failure analysis, Strong foundations in classical computer vision and image processing, with experience including tools such as OpenCV, NumPy, or equivalent libraries, Fluency in Python and a modern deep-learning framework such as PyTorch, Experience designing evaluations that support real production decisions, including metric selection, operating thresholds, calibration, dataset construction, slicing, leakage prevention, and regression analysis, The ability to write maintainable research and production-quality code, A pragmatic, applied-research mindset: rigorous about experiments, but able to determine when a result is ready to ship and when further optimization has diminishing value, Strong written communication.

You define the problem, requirements, assumptions, experiments, success criteria, and conclusions clearly, A collaborative operating style: you work independently without becoming isolated, and you actively share context and knowledge with others, engaging constructively with constraints from neighboring teams rather than treating them as obstacles.

We are not looking for a lone wolf, however brilliant, An "in-the-driver's-seat" operating style: you take ownership of problems end-to-end, drive your work forward without waiting for direction, and stand behind your decisions once they are in production, Direct experience with biometric verification/identification; margin-based metric learning losses and their failure modes; presentation attack and liveness detection; or adversarial evaluation of ML systems, Experience with edge optimization and on-device deployment of ML models.

Quantization, pruning, distillation, kernel-level optimization, deployment to mobile NPUs, embedded GPUs, microcontrollers, or other constrained targets, Hands-on experience with Rust for high-performance code paths, and the disposition to optimize for speed rather than treat it as someone else's problem, A background in sensors, imaging, computational photography, or camera ISPs, Familiarity with privacy-preserving computation or machine learning systems that interact with secure multi-party computation