Building safe generative AI with human touch
Subhabrata (Subho) Mukherjee, PhD
Co-Founder & Chief Scientific Officer, Hippocratic AI
At Hippocratic AI, I lead the technology organization building safe generative conversational AI for healthcare. As an AI and technology executive, I spearhead next-generation multimodal AI models that align to human reasoning, emotion, and safety. My executive scope spans building AI platforms and leading organizations across speech, large language models, and efficient inference for frontier-scale systems. Hippocratic AI has received $404 million in funding at a $3.5B valuation.
Prior to this, I led large-scale foundation-model initiatives as a Principal Researcher at Microsoft Research, with a decade of AI work across Microsoft, Amazon, IBM, and Google and 100+ publications. I earned my PhD summa cum laude from the Max Planck Institute for Informatics and was the 2018 SIGKDD dissertation runner-up.
01 / Building the organization
Building the organization
behind safe AI at scale.
Every deployment creates signals that improve the system; every research result has a path to improving real human - AI interaction.
Polaris
A safety-first system—not a single model.
A stateful primary model holds natural, long-form voice conversations while specialist models for medication safety, labs, compliance, escalation, and memory work in parallel. A shared orchestration layer coordinates the system in real time.
In-house built, state-of-the-art ASR and TTS make conversations accurate, natural, and clinically fluent.
An in-house inference stack serves a 700B+ primary LLM at real-time voice latency.
Speech
In-house ASR · in-house TTS · real-time voice
Frontier models
Post-training · large-scale reinforcement learning · recursive self-improvement
Alignment
Reasoning · emotion · safety
Inference
Quantization · kernels · decoding · network & communication
Clinical platform
Orchestration · evaluation
Applied research
Paper → product → real patient
Backed by leading investors including Andreessen Horowitz, General Catalyst, Kleiner Perkins, NVIDIA Ventures, Alphabet CapitalG, Avenir, and Premji Invest.
02 / Speaking & press
The public voice
of the technology.
Keynotes, features, and invited talks on safe agentic AI, alignment, and building systems that scale.
The Agentic AI Advantage
On infusing generative conversational AI with genuine human touch—and the engineering behind agents that understand, reason, and connect safely with patients.
03 / Research
Research
that ships.
100+ publications and patents across a decade at Microsoft Research, Amazon, IBM, and Google—now compounding inside a production system.
Orca
Progressive learning from complex explanation traces helped define a new era of open-model post-training.
650+ citationsRead paper ↗Polaris
A safety-focused LLM constellation built for real-time voice conversations in healthcare.
250M+ interactionsRead paper ↗Human–AI interaction
Beyond what real patient conversations teach, this work shows the system building blocks: in-house ASR, TTS, inference, and natural human–AI interaction alignment—and how research is translated into a real-world production system.
NeurIPS keynoteRead paper ↗HEART
A unified benchmark for assessing humans and language models in emotional-support dialogue—connecting rigorous evaluation with the empathy, reasoning, and safety required in real conversations.
Safety & alignment
Red-teaming, reversible fine-tuning, real-world clinical evaluation, and benchmarks for emotional support.
Efficient inference
KV-cache compression, MoE megakernels, query routing, and efficient reasoning at frontier scale.
Real-time voice
Parallel speech recognition, streaming decoding, and low-latency synthesis for natural conversation.
Agents & retrieval
Graph-grounded generation and trajectory-aware evaluation for agents working in the real world.
Research protected.
Systems deployed.
Seven US patents and applications spanning conversational safety, speech, memory, and model learning.
Low-latency conversational AI with a parallelized in-depth analysis feedback loop
↗US 12,367,971Low-latency analysis feedback for safety-focused conversational AI
↗US 12,597,512Real-time use of multiple parallel ASR modules with selection and fusion
↗US App. 19/400,096Multi-call memory for longitudinal AI care conversations
↗US 2025/0094827Producing a reduced-size model by explanation tuning
↗US 12,073,326Joint learning from explicit and inferred labels
↗US 11,797,755Unsupervised annotation generation for natural-language understanding
↗