Vibhor Kumar is a computer scientist and AI engineer specializing in trust and safety at scale. After building machine-learning systems at Visa, Imbellus, Tinder, and Cinder, he is developing classifai.dev, a classification platform for developers and AI agents, and open-source safeguards for autonomous software.
Kumar studied computer science at Caltech, concentrating on machine learning, and later earned an MBA from Quantic. His research has investigated linguistic structures in distributed memory and curriculum learning for reinforcement-learning agents. At Visa, he applied generative models to transaction fraud and merchant recommendations; at Imbellus, he developed automatically generated assessment tasks and reinforcement-learning agents.
At Tinder, Kumar built adaptive defenses against spam, scams, impersonation, and harassment. His 2024 AI Engineer presentation outlined how fine-tuned open language models could identify semantically complex abuse while remaining economical enough for high-volume production. He subsequently became Cinder’s founding AI/ML engineer, developing harmful-content detection and agentic moderation, before turning to independent products.
- Adversarial abuse detection: Fraudsters evade static rules and similarity matching by changing language, imagery, and account behavior. Kumar combines semantic classification with harder-to-fabricate signals, including network and account metadata, to recognize coordinated activity as tactics evolve.
- Human-AI hybrid training data: Large models identify potential violations within real platform data; human specialists then correct labels and resolve ambiguous policy decisions. Grounding datasets in actual behavior avoids the distortions of exclusively synthetic examples.
- Multi-adapter inference: Parameter-efficient LoRA fine-tuning produces specialized detectors that share one base model. Using LoRAX, developed by Predibase, teams can serve multiple harm-specific adapters efficiently and combine them with filtering, model cascades, and retraining.
- Classification as developer infrastructure: classifai.dev extends these techniques into multimodal classification, moderation, request routing, feedback-driven improvement, and agent integrations.
- Agent execution governance: Kumar’s AEGIS project evaluates agent actions against predicted risk, tool sensitivity, and previous user preferences, allowing routine operations while escalating consequential ones for confirmation. His ClaimValidator project applies related scrutiny to whether cited sources substantiate online claims.
