The LLM Data Company provides purpose‑trained large language models for clinical decision support, delivering evidence‑based, safety‑aware responses. Its Kos‑1 Lite model, a ~100 B parameter mixture‑of‑experts, achieves state‑of‑the‑art performance on HealthBench Hard and is accessed via a HIPAA‑compatible inference API with built‑in deferral and calibrated uncertainty for hospitals and medical AI platforms.
Funding
$500K raised to dateRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Product
Problem
General-purpose large language models exhibit excessive sycophancy, verbosity, and insufficient grounding for high‑stakes domains such as clinical decision support, making them unreliable for medical reasoning tasks that demand precise, deferrable, and safety‑aware responses.
Solution
The LLM Data Company builds “frontier” models that are purpose‑trained for underserved, high‑risk domains. By applying hand‑crafted post‑training curricula and reinforcement‑learning objectives that penalize unwarranted agreement, the company produces models that prioritize clinical accuracy, calibrated uncertainty, and a compassionate bedside manner. Their flagship Kos‑1 Lite model achieves state‑of‑the‑art performance on the HealthBench Hard benchmark while operating at a fraction of the serving cost of trillion‑parameter systems. The models are delivered via a cloud‑hosted inference API that integrates with existing health IT stacks, enabling real‑time, evidence‑based assistance without the need for extensive prompt engineering.
Target Audience
Primary customers are hospitals, health systems, and medical AI platforms that require reliable, domain‑specific language models for clinical decision support, as well as research institutions developing advanced medical reasoning applications.
Features
- ~100 B parameter mixture‑of‑experts (MoE) architecture trained with Miles (SGLang/Megatron) for scalable inference.
- Hand‑crafted post‑training curricula that explicitly trade off coding proficiency for reduced sycophancy in clinical contexts.
- Reinforcement‑learning reward function with soft‑overlong penalties and quality‑threshold bonuses to encourage thorough internal reasoning while producing concise user‑facing answers.
- HealthBench Hard SOTA score of 46.6 % and competitive performance against GPT‑5 at significantly lower compute cost.
- Built‑in deferral and triage mechanisms that recognize out‑of‑scope queries and suggest human review.
- Secure, HIPAA‑compatible API with role‑based access control and end‑to‑end encryption for integration into EHR and CDSS platforms.
- Comprehensive evaluation pipeline including physician review, safety testing, and longitudinal benchmarking.