Tanishq Abraham

Tanishq Mathew Abraham is CEO and co-founder of Sophont, a company developing open multimodal foundation models for medicine, alongside CTO and co-founder Paul Scotti. Previously a Research Director at Stability AI, Abraham founded and led MedARC, an open medical AI research organization. His work includes collaborative research on reconstructing visual perception from brain activity and a public argument for medical models that combine information across patient-data modalities.[3][2][6][5]

On March 31, 2025, Abraham and Scotti announced Sophont in their jointly authored Medical AI Manifesto. They proposed a lab combining medical data and clinical expertise with large-scale model-training capabilities, while translating open research into practical applications. Their description was explicit: “We are a research and translation company.” Abraham's account credits his younger sister, Tiara Abraham, with suggesting the company name.[5]

The founders argue that medical-specific training need not mean narrowly specialized architectures. Generalist language models can perform well on clinical text, they acknowledge, but three-dimensional CT scans and gigapixel pathology slides require different inputs and training data. They favor methods that scale with compute and data, and propose combining specialized models into a holistic system processing lab tests, imaging, ECG time series, surgical video, and clinical notes. This is their research agenda, rather than a claim that such a universal model is already complete.[5]

Abraham and Scotti also argue for open-source medical AI because available weights and training information can support auditing and interpretability. Local deployment would let healthcare providers keep patient information on their own premises, while retaining a fixed model version could avoid unexpected behavior changes from service updates. Sophont's stated approach pairs open weights with public evaluation; its first commercial product remains in development.[5][3]

Before the Sophont announcement, Abraham's work spanned digital pathology, open-source teaching, and research collaboration at Stability AI and MedARC.[2][5]

Abraham completed his biomedical engineering PhD at UC Davis in spring 2023. He describes researching deep learning—especially generative AI—for novel microscopy techniques in digital pathology under Richard Levenson. His personal account also identifies him as a founding-team member of DALL-E mini and a frequent fastai contributor, with teaching roles as a Part 1 teaching assistant and Part 2 instructor for the 2022 fast.ai course.[4][2]

At Stability AI, Abraham founded MedARC to pursue open research on medical foundation models. He also describes helping manage the company's external compute cluster and relationships with collaborators receiving compute for generative AI research. In the Sophont manifesto, he and Scotti recall expanding MedARC projects through external collaborators around the world.[2][5]

Abraham coauthored MindEye2, published at ICML in 2024, with Paul Steven Scotti and nine other collaborators. The manifesto credits Scotti with leading the team behind the MindEye publications. MindEye2 addressed the data burden of models trained separately for each person: the authors reported high-quality reconstructions of visual perception using one hour of a new subject's fMRI training data, after pretraining across seven subjects.[6][5]

The team's method linearly aligns brain data into a shared-subject latent space, then applies a shared nonlinear mapping into CLIP image space. Stable Diffusion XL, fine-tuned to accept CLIP latents instead of text, converts those representations into pixels. The reported one-hour result concerns adaptation to a new subject after cross-subject pretraining and reconstruction of visual perception.[6]