Discovery
Turn a vague research need into a bounded product decision: should a care team run a pilot?
SUVANÉRESEARCH Research prototype
Case study · Building in public
From research fragments to a defensible product decision.
SUVANÉ Research is a human-in-the-loop AI implementation case: discovery, data modeling, retrieval, structured synthesis, evaluation, deployment, and operational safeguards.
ProblemResearch is abundant. Decision context is missing.
UserHealth product, innovation, and care teams.
OutcomeA traceable pilot recommendation—not another summary.
System workflow
Normalize
Map papers and reports into one evidence schema: population, intervention, outcomes, strength, limits, and implementation meaning.
Retrieve
Search source chunks and records together, then require every synthesis claim to remain traceable.
Synthesize
Generate a fixed decision brief with explicit uncertainty and a refusal path when evidence is insufficient.
Evaluate
Test retrieval coverage, citation correctness, unsupported claims, answer consistency, and usefulness to target professionals.
Review
Keep evidence ratings and publishing under human control; log source, prompt, model, output, and approval state.
Architecture direction
A modular monolith before a platform.
The live decision surface now calls a FastAPI service with deterministic retrieval, structured Pydantic contracts, verified evidence records, and an explicit insufficient-evidence path. Postgres/pgvector, ingestion workers, and reviewer state remain later milestones.
Verified sources→Deterministic retrieval→Decision brief→Source traceability→Human judgment
Evaluation plan
The demo is not complete when it produces an answer.
Retrieval
Did the system find the evidence needed to answer the question?
Grounding
Does every material claim match and cite a retrieved source?
Calibration
Does confidence fall when evidence quality or coverage is weak?
Decision utility
Can a professional identify the next responsible action?