Community
Open Source
Tools from our applied-AI work, released as they mature. A common thread runs through all of them: every pipeline is designed to run on local GPU infrastructure — vLLM inference, multi-GPU scaling, reproducible caching — with no cloud dependency.
-
trialmesh
Clinical-trial matching on local GPUs
A modular system that semantically matches cancer patients to relevant clinical trials: LLaMA-summarized patient and trial records, biomedical embeddings (SapBERT, BioClinicalBERT, and others), FAISS vector search, and multi-stage LLM reasoning with explainable match decisions. Built to run entirely on-premises via vLLM with distributed multi-GPU support.
- Python
- vLLM
- FAISS
- Apache-2.0
-
crater
Temporal entity resolution in clinical records
Clinical Record Analysis for Temporal Entity Resolution: generates synthetic MIMIC-IV-style patient cohorts with a clinical state machine, then uses wave-batched vLLM inference to resolve whether condition mentions across sequential encounters refer to the same instance. Ships with a full multi-GPU benchmarking harness (crater-bench) and evaluation notebooks.
- Python
- vLLM
- Llama 4 / MedGemma
- MIT
-
chromatic_orchestrator
Data-parallel vision classification pipeline
A production-oriented three-stage vision pipeline — classify, score, summarize — on multimodal Llama models. Hash-based caching with rank isolation, configurable tensor parallelism, and file-based prompt customization; scales from a single GPU to data-parallel clusters processing 1,000+ images per minute.
- Python
- vLLM
- Multimodal LLMs
- Apache-2.0
-
TrialGPT (rework fork)
Reworked patient–trial matching research code
An extensive rework fork of NCBI’s TrialGPT — the Nature Communications framework for zero-shot patient-to-trial matching with criterion-level eligibility predictions and hybrid BM25 + MedCPT retrieval.
- Python
- LLM eligibility matching
- Fork of ncbi-nlp/TrialGPT
Something Here Fit Your Problem?
These pipelines generalize — trial matching, record analysis, and vision classification are patterns, not one-offs. Talk to us about adapting them to your data, on your hardware.
Contact Us