Vikhyat “Vik” Korrapati is the co-founder and chief technology officer of Moondream, which builds compact, open vision-language models for practical image understanding. His central engineering conviction is that visual intelligence must operate close to where images are captured, with predictable costs, low latency, and safeguards suited to the application.
Korrapati previously built an anime recommendation engine for Hummingbird, later known as Kitsu, and spent approximately nine years working on distributed systems at Amazon Web Services. Moondream grew out of his subsequent attempt to develop an automated browser-testing agent: existing vision systems were too slow, expensive, and restrictive to reliably interpret ordinary browser screens. His smaller alternative attracted developers facing similar problems, turning a testing side project into a broader computer-vision company co-founded with Jay Allen. In October 2024, Korrapati announced $4.5 million in funding backed by Felicis, Microsoft’s M12 GitHub Fund, and Ascend.
The original Moondream open-source project combined Google’s pretrained SigLIP vision encoder with Microsoft’s Phi-1.5 language model to support image captioning, visual question answering, object localization, and counting. Korrapati prioritizes adapting proven components over expensive training from scratch and improving useful visual tasks over demonstrations of general intelligence.
- Training data over architectural novelty. Korrapati argues that small models fail when synthetic examples misrepresent actual user behavior: indiscriminate distillation can encourage hallucinations, repetitive distractors can teach superficial patterns, and polished prompts can leave models unprepared for typos. His LNQA visual question-answering dataset addresses those problems with spatially grounded descriptions, irrelevant questions, and openly licensed material across approximately 303,000 images and 1.5 million question-answer pairs.
- Deployable visual intelligence. For video analysis, robotics, and monitoring, Korrapati treats inference cost, responsiveness, and privacy as product requirements. Moondream 3 uses a nine-billion-parameter mixture-of-experts architecture with approximately two billion active parameters per token. His Kestrel inference engine adds asynchronous micro-batching, streaming, efficient caching, and optimized NVIDIA CUDA and Apple Metal execution.
- Developer-controlled computer vision. Natural-language prompts let developers describe what they need without collecting labels and training custom detectors for each task. Korrapati also favors application-layer safety controls, allowing developers to calibrate protections for their specific use case instead of inheriting blanket model restrictions.
His more recent RefCOCO-M visual-grounding dataset improves segmentation masks and removes problematic prompts, extending his focus on practical reliability from model training and deployment into evaluation.
