Less manual work.
More meaningful
insight.

The problem

Find reliable answers across multiple uploaded documents while identifying which source supports each claim.

What I built and tested

Document ingestion, vector and keyword retrieval, reranking, conversational follow-ups, citation validation, and bounded self-correction. The system avoids unsupported answers when evidence is insufficient.

Evaluation

71 automated tests; 41 manually labeled retrieval questions. At top-2, recall was 0.88 and mean reciprocal rank (MRR) was 0.92 on the project benchmark, not a universal accuracy claim.

Development approach

Developed with AI coding assistance and hands-on requirements definition, iteration, validation and testing.

The problem

Make database questions easier to explore without letting generated SQL silently modify data.

The approach

Schema-aware query generation, SQLGlot validation, explicit approval tickets, database-level read-only controls, saved queries and visual outputs.

Verification

The local SQLite demo and automated mocks were tested, with 63 automated tests covering query generation, security and application behavior. Live multi-provider generation depends on configured API access.

Development approach

Developed with AI coding assistance and hands-on requirements definition, iteration, validation and testing.

The problem

Move data between different systems with traceability, consistent validation and safe retry behavior.

The approach

Built extraction and loading for files, PostgreSQL and REST APIs, optional Gemini-assisted schema mapping, batch processing, checkpoints, per-record outcomes, audit logs and reconciliation.

Measured test

A 10,000-record synthetic input processed in 10.48 seconds: 9,933 inserted, 10 skipped and 57 rejected. All 9,933 inserts reconciled; the rerun introduced zero duplicates. These results are from a local acceptance test, not a public hosted benchmark.

Development approach

Developed with AI coding assistance and hands-on requirements definition, iteration, validation and testing.

The challenge

Booking data needed cleaning and structure to support more efficient processing.

The approach

Cleaned records, introduced validation, and improved the database schema and indexing. Added SQL quality checks, scheduled backups and recovery testing.

The result

An 83% reduction in booking-processing time, as reported in my résumé.

The challenge

A daily retrieval process needed to account for missing meter readings.

The approach

Scheduled retrieval for 9:00 AM. If no reading existed after 8:50 AM, the procedure selected the maximum reading before 8:50 AM. If unavailable, it checked the previous day’s raw data, then its backup data.

The deliverable

A repeatable retrieval workflow with defined fallback paths.

The focus

Bring service-ticket activity and team workload into a regularly refreshed view.

The deliverable

Dashboards tracking cleaning and maintenance volumes, response times and workload, refreshed hourly.

The approach

Connected date-based reporting in Power Apps to SharePoint records and automated PDF generation.

Related delivery

Also built a task-progress dashboard covering overdue items, pending actions and critical delays. Resolved production workflow failures through data restructuring and backend cleanup.

The impact

The finance task-progress dashboard reduced manual reporting time by 40%, as documented in my résumé.