CLARA AGENT RELIABILITY

A review flow
you can follow.

A local, rules-based demo for tracing how a product query moves through catalogue retrieval, claim and risk checks, and a doctor-review hold.

Explore the project Watch the recorded demo

20catalogue entries

0model calls

Localrule-based flow

Screenshot of the local Clara review workspace with sample query chips, a blank query area, and the review-flow overview — Local browser workspaceCaptured from the running demo

THE APPROACH

Make each step visible.

Clara demonstrates a small control layer around a manually maintained product catalogue. It is a project demo, not a healthcare service.

Route

Rule-based checks classify the query and flag selected high-risk terms.

Retrieve

Keyword matching ranks products from the checked-in JSON catalogue.

Ground and check

Catalogue claims, risk factors, and contraindications are surfaced with each result.

Hold and trace

When a rule requires review, the result shows the hold and the recorded decision path.

FROM THE LOCAL DEMO

A decision with its context.

These existing screenshots show the local interface and a sample review result. They are static images; this presentation does not run that workflow.

Existing screenshot of a sample query held for doctor review, with catalogue claims, risk factors, and the decision trace — Sample review outputIllustrative rules-based demo result · not medical advice

Existing narrow-screen screenshot of the local Clara workspace

RESPONSIVE LOCAL UI

The workspace also adapts to a narrow screen.

The screenshot is from the existing local interface. The static presentation remains read-only at every viewport.

WANT TO RUN IT?

Start the interactive demo locally.

The Python server provides the query-review and optional feedback endpoints on your own machine. This hosted presentation does not provide either endpoint.

From the repository root · Python 3.9+

python3 -m src.web_app --host 127.0.0.1 --port 8765

Read the full setup instructions

SCOPE AND LIMITS

A prototype with clear boundaries.

Deterministic keyword rules; no language-model calls or semantic retrieval.
Small, local catalogue with manually maintained product data.
No clinician integration, live product database, or remote service.
Not for real medical use. The sample output is only a demonstration of the project rules.