Reject defects without losing evidence
Ordered rules normalize accepted values, preserve safe exception records, and reconcile every deliberately injected defect.
Featured evidenceSynthetic dataHuman-reviewed AI
AI-Assisted Data Quality & Reporting Workflow
I designed and implemented a reproducible Python and SQL workflow that validates operational data, reconciles metrics, produces Excel and HTML reports, and controls AI-generated summaries through automated claim checks and separate human approval.
✓ 900 synthetic records generated
✓ 840 accepted · 60 rejected
✓ 10 SQL metrics reconciled
✓ Excel + HTML reports produced
✓ Structured AI draft validated
✓ Human decision: APPROVED900Synthetic source records
66Injected defect events reconciled
10Approved SQL metrics
93Automated project tests passed
Operational problem
Turn imperfect records into trustworthy reporting
Reporting workflows need clear validation, traceable metrics, usable outputs, and safeguards when AI assists with narrative drafting.
Separate approved metrics from AI wording
SQL defines authoritative values. AI receives aggregate metrics only and cannot calculate, approve, or publish its own claims.
Keep useful outputs available without AI
Deterministic JSON, Excel, HTML, and accessible charts remain complete when a provider fails or a draft is rejected.
Require automated controls and a person
Schema, numeric, semantic, privacy, and limitation checks run before an append-only human review decision.
Implementation
A controlled path from raw data to reviewed insight
- GenerateFixed seed and defect manifest
- ValidateNormalize, reject, reconcile
- AnalyzeSQLite and versioned SQL
- ReportJSON, Excel, HTML, charts
- Review AIClaims checked, then approved
Published evidence
Sanitized artifacts recruiters can inspect
These files preserve aggregate results and review evidence while excluding credentials, account details, provider payloads, local paths, rejected drafts, and row-level records.
- HTTP 200 live Responses API demonstration with a fixed model snapshot
- 3,921 total tokens recorded for the passing attempt
- All automated claim controls passed before human review
- Ten human-review checklist items passed and final state recorded as approved
- Temporary restricted credential revoked after verification
- Exact artifact hashes enforced by contract tests
verification-summary.json
approved-narrative.json
review-receipt.json
evidence-manifest.jsonEvidence boundary
A professional demonstration—not a production claim
This is a self-directed portfolio project using fictional data. The live API call was a bounded local demonstration, not a client deployment, enterprise workload, model-development result, or measured organizational impact.
Related capabilities
Data analytics, AI controls, systems thinking, and automation
The project demonstrates how I combine technical implementation with evidence, documentation, safe boundaries, and practical reporting.