SUBHABRATA (SUBHO) MUKHERJEE
Hi, I am Subho. I am the co-founder and chief scientist at Hippocratic AI, a new startup in Generative AI and Healthcare. Hippocratic, featured in Fortune 50 AI Innovators (2023) and Nature Portfolio Medicine, is building a safety focused large language model (LLM) for the healthcare industry. We are leveraging generative AI to revolutionize healthcare access with AI-powered safe conversational agents. Hippocratic AI backed by General Catalyst, Andreessen Horowitz, Premji Invest and others raised $120M in funding at $500M valuation.
I head AI teams building the next generation of foundation model training and conversational alignment to human reasoning, preferences and clinical safety.
Contact Me:
RECENT NEWS
Transforming Generative AI and Healthcare
Technical Report – Polaris 1T+ LLM Constellation for Healthcare
[Mar, 2024] We are excited to share the technical and clinical considerations that went into building Polaris in our 53-page technical report available now in ArXiv. Key highlights:
1. Polaris is the first safety-focused LLM for healthcare geared for real-time patient-AI voice conversations.
2. Polaris is a 1T+ parameter constellation system composed of several 70B-100B parameter LLMs trained as co-operative agents: a stateful primary agent driving an engaging conversation and several specialist agents focused on tasks performed by nurses to increase safety and reduce hallucinations.
3. Polaris performs on par with U.S. licensed human nurses on aggregate across dimensions such as medical safety, clinical readiness, conversational quality, patient education, and bedside manner.
4. For Phase 2 testing, we recruited 1100 U.S. licensed nurses and 130 U.S. licensed doctors for end-to-end conversational evaluation— the most extensive evaluation performed to date for any healthcare LLM.
5. Polaris enables Rachel to have those amazing conversations in the GTC24 demo
Hippocratic AI Raises $53M At a $500M Valuation
[Mar, 2024] We announced the close of a $53 million Series A funding round at a $500 million valuation, bringing total funding to $120 million. We also released its first product for phase three safety testing: a staffing marketplace for healthcare where health systems, payors, and others can “hire” generative AI agents that complete low-risk, non-diagnostic, patient-facing healthcare tasks to help solve the massive shortage of healthcare nurses, social workers, and nutritionists in the US and worldwide. The new capital will be used to help fund these safety tests and further develop the product.
The round was co-led by Premji Invest and General Catalyst with participation from SV Angel and Memorial Hermann Health System as well as existing investors Andreessen Horowitz (a16z) Bio + Health, Cincinnati Children’s, WellSpan Health, and Universal Health Services (UHS).
Hippocratic AI – Nvidia Partnership for Empathy Inference
[Mar,2024] We announced a collaboration with NVIDIA to develop empathetic AI healthcare agents — powered by the NVIDIA AI platform – enabling super-low-latency conversational interactions. User tests repeatedly show that super low latency voice interaction is required for patients to build an emotional connection naturally. Since LLMs run on inference engines, we term this low latency inference: “Empathy Inference.”
The AI healthcare agents are built on our safety-focused large language model (LLM), the first designed specifically for healthcare. Health systems, payors, digital health companies, and pharma deploy our healthcare agents to augment their human staff and complete low risk, non-diagnostic, patient facing tasks over the phone.
Rachel at Nvidia GTC 2024
[Mar, 2024] The first sneak peek into Rachel, the first real-time conversational healthcare agent — what Jensen Huang introduced to the world as a digital human at NVIDIA GTC 2024 conference.
Featured in Fortune 50 AI Innovators
[Nov, 2023] Featured in the Fortune 50 AI Innovators 2023, Hippocratic is transforming the field of Generative AI and Healthcare. Given the massive shortage in healthcare staffing, we are leveraging generative AI to fill in the void with AI-powered safe conversational agents. For instance, you can talk to our virtual nurses who can follow-up on your health conditions and care plans, schedule follow-up appointments, review medication issues and help manage chronic conditions for non-diagnostic purposes.
Echoing the sentiment from the blurb put out by Fortune: “One thing we know for sure: The work these companies are doing won’t just shape the future of AI, it will shape the world we all live in.”
Featured in Nature Medicine
[Nov, 2023] With our safety-first LLM, we are looking to fill existing health care staffing gaps (“superstaffing”) and enable interventions by decreasing the cost of non-diagnostic services.
We train our own model using proprietary data including clinical care plans, regulations, medical manuals, etc. and teach it medical reasoning. We perform extensive alignment to teach the model how to speak like healthcare workers using conversations between U.S. licensed nurses and patient actors. We conduct a unique RLHF process using healthcare professionals to train and evaluate the model on fine-grained aspects including knowledge, bedside manners, conversation style and clinical task completion.
Launching Physician Advisory Council
[Oct, 2023] Given our safety-first focus, we formed the Physician Advisory Council comprising of expert physicians from leading US hospitals, health systems and digital health companies, who will play a crucial role in guiding the development of our technology and ensuring that it is ready for safe deployment.
Early Access Partnership Program
[Dec, 2023] Early Access Partnership Program brings together several healthcare industry innovators to partner with us in the development of our technology and validate priority use cases.
Healthcare organizations will obtain early access to our products and play an integral role in contributing to our LLM’s development and safety to revolutionize healthcare.
Recent Publications
Orca: Progressive Learning From Complex Explanation Traces of GPT-4
[Jun, 2023] Orca is a new AI model that can learn from explanations, step-by-step thought processes, and other complex instructions. Despite being a small model with 13 billion parameters, Orca demonstrates competitive performance in professional and academic examinations such as SAT, LSAT, GRE, and GMAT and reaches parity with ChatGPT on Big-Bench Hard (BBH).
Teaching Language Models to Hallucinate Less with Synthetic Tasks
[Nov, 2023] SynTra shows that reducing hallucination on synthetic tasks can also reduce hallucination on real-world downstream tasks. SynTra optimizes LLM’s on synthetic tasks where hallucinations are easy to elicit and measure and transfers to realistic, hard-to-optimize tasks using only a synthetic retrieval task for supervision.
Skipdecode: Autoregressive Skip Decoding with Batching and Caching for Efficient LLM Inference
[Jul, 2023] SkipDecode enables faster text generation for Autoregressive Language Models using reduced computation without applying the full computation graph to each token. Unliked prior early exit methods, SkipDecode develops efficient batch inferencing and KV caching with 2x to 5x inference speedups with negligible regression across a variety of tasks.
AutoMoE: Heterogeneous Mixture-of-Experts with Adaptive Computation for Efficient Neural Machine Translation
[Jul, 2023] AutoMoE is a framework for automatically designing heterogeneous MoE’s (e.g., how many experts? where to place them? what should be their sizes?) under computational constraints. AutoMoE leverages Neural Architecture Search (NAS) to obtain efficient sparse MoE sub-transformers with 4x inference speedup and FLOPs reduction over manually designed Transformers. AutoMoE allows adaptive compute–where different computation is used for different tokens in input.
AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning
[Nov, 2022] AdaMix is a general PEFT method that tunes a mixture of adaptation modules like a mixture of Houlsby adapters or a mixture of low rank decomposition matrices like LoRA. By only tuning 0.1-0.2% of LLM parameters, AdaMix outperforms SOTA parameter-efficient fine-tuning and full model fine-tuning for both NLU and NLG tasks.
ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models
[May, 2023] Augmented Language Models (ALMs) blend the reasoning capabilities of LLMs with tools that allow for knowledge retrieval and action execution. ReWOO (Reasoning WithOut Observation) is a modular paradigm that detaches the reasoning process from external observations, thereby significantly reducing token consumption with 5x token efficiency and accuracy improvements. We offload reasoning ability from 175B GPT3.5 into 7B LLaMA, demonstrating the ALM potential.
Subhabrata (Subho) Mukherjee 