DATA & AI · BY GUSTAVO

Project
Summary

Explore the business question and the implementation behind this project.

Explore the summary

WHY THIS MATTERS

Monthly Totals That Can Be Reconciled

THE PLATFORM, IN NUMBERS

Trace the Data. Inspect the Result.

Explore the documented validation run.
Every figure has a source.

01 / RECORD ACCOUNTABILITY

Every Record Has a Destination

Known-answer fixture

69/ 69

input rows reconciled

One exact duplicate is removed. Eight invalid claims are quarantined. The remaining 60 claims enter the validated dataset.

60 accepted8 rejected1 duplicate

RECORD MAP1 tile = 1 input row

69 rows accounted for. No unexplained row loss.

The local DuckDB fixture is separate from the generator’s 50,100 claims and 1,200 providers. Validation record

02 / PAYMENT STATUS

Accepted Does Not Always Mean Paid

Only paid claims contribute to monthly payment totals. Denied claims stay in the validated dataset.

PAID CLAIMS4066.7% of accepted claims

40 paid claims contribute to the report. The other 20 accepted claims are denied.

Status counts from the fixed verification dataset. Percentages use 60 accepted claims as the denominator. View source

03 / MONTHLY REPORTING

40 Paid Claims. Four Report Rows.

Two specialties across two months. Inspect the counts that roll up into the monthly report.

40paid claims · both months

CardiologyPrimary carePaid claims

November: 19 · December: 21. All four report rows are checked against known answers.

Counts and payment totals are validated by the harness. This chart shows counts from the development fixture. Verification harness

04 / AUTHENTICATED FASTAPI SERVICE

An API for Claim Summaries and Document Q&A

51API / RAG tests passed
recorded local suite

I built the API layer to expose claim summaries, retrieve reference documents and return answers with checked citations.

REQUESTClaim questionReference documents

API-KEY AUTHENTICATIONFastAPIRetrieve · summarize · cite

RESPONSECited answerEvidence can be inspected

HOW IT WORKS

Gemini embeddings and pgvector retrieve reference passages. The API summarizes the evidence and checks citations before returning an answer.

IMPLEMENTED CONTROLS

API-key accessVector retrievalCitation checks

The 51-test local suite mocks or intercepts provider calls. Live Gemini answer quality is a separate evaluation. The diagram describes the implemented service; it is not a live API call. API architecture

05 / PUBLICATION SAFEGUARDS

A Failed Quality Gate Protects Published Data

16 data checks · 8 dbt models

Select a documented test scenario to see what happens to the build and the previously published results.

VERIFIED OUTCOMERepeatability confirmed

BUILDPasses

PUBLISHED RESULTSUnchanged

Accepted claims, monthly reports and the quality summary match the first successful build.

Recorded local DuckDB checks compare complete outputs before and after each scenario. These interactions display documented results. Validation record

EXECUTIVE OVERVIEW

My Contribution Across the Platform

THE WORKFLOW

What Is Verified, and What Comes Next

Explore on GitHub

Reviewed 22 October 2026 against the project README and linked evaluation reports. Figures describe the documented dataset or test run.