Descriptions:
Charlie Wood, global architect at Acten, and William Imoh presented at AI Dev 26 in San Francisco showcasing a healthcare AI agent designed to reduce readmission risk and automate pre-visit chart preparation. The clinical context is concrete: clinicians currently spend roughly 45 minutes per patient on manual chart prep across fragmented systems, and approximately 15% of patients are readmitted within 30 days — a period during which insurance reimbursement is denied.
The centerpiece of the technical stack is VectorAIDB, a local-first vector database that Acten launched at the event. Designed for regulated industries where patient data cannot leave the network, it supports real-time indexing, sub-second retrieval, and a developer-first API. Benchmarks presented showed a 3–7x query-per-second improvement over competitors at 1 million vectors, scaling further at 10 million vectors, with 0.988 P95 recall versus approximately 0.99 for competing base configurations — a modest recall tradeoff for substantially better throughput under load.
The agent architecture uses two sequential stages: a context-gathering agent retrieves patient-specific records from the vector database by patient ID, and a risk analysis agent synthesizes deterioration signals and medication conflicts from that retrieved context. The production FastAPI implementation, shown in approximately 400 lines of Python, covers prompt injection mitigation, data validation, semantic chunking of discharge notes, and filterable metadata predicates — patient ID, encounter date, length of stay, and source — enabling precise retrieval across large patient datasets without exposing PII outside the network perimeter.
📺 Source: DeepLearningAI · Published May 20, 2026
🏷️ Format: Workflow Case Study
