Captured from the live deployment, not a mockup
01 · The problem
Clinical text cannot be handed to a hosted model as-is, so every identifier is masked before the request leaves the service, enforced across a 38-endpoint API rather than at one choke point.
02 · How it works
- 01
Benchmarked nine Amazon Bedrock models on latency and multilingual quality and selected Nova Pro at 605 ms, setting maxTokens on every call to avoid quota-reservation throttling.
- 02
Matched prescriptions to Jan Aushadhi generics with substitution-safety warnings, surfacing savings up to 93.7%, and simplified clinical reports into 15 Indian languages.
- 03
Masked every identifier before it reaches the model across a 38-endpoint API, so a hosted model never sees patient-identifying text.
- 04
Benchmarked nine Bedrock models on latency and multilingual quality to select Nova Pro at 605 ms, and set maxTokens explicitly to avoid quota-reservation throttling.
03 · What it cost, and what it returned
Built a clinical assistant that masks every patient identifier before a request can reach a model: a 38-endpoint FastAPI API and Next.js front end over 11 DynamoDB tables, reports in encrypted S3 behind 24-hour presigned URLs, and IAM scoped to each table, bucket and model.