OnFinance AI

AI-Powered Bulk Upload Pipeline

The Challenge

Compliance teams were drowning in manual data entry. Creating 10,000 regulatory tasks took 40+ hours per quarter, with a 12% error rate causing compliance violations. Customers demanded bulk upload, but existing architecture couldn't handle it.

The Solution

Designed and built a complete bulk ingestion system from ground up: multi-step wizard frontend, FastAPI backend with 9 RESTful endpoints, AI-powered column mapping using GPT-4, and background workers for processing. System handles Excel/CSV files with thousands of rows, automatically maps columns to schema, and creates validated tasks in under 60 seconds.

Technical Highlights

Code Sample: Batch Write with Graceful Degradation

async def create_tasks_batch(tasks: List[Dict]):
    CHUNK_SIZE = 1000
    inserted_ids = []

    for chunk in chunks(tasks, CHUNK_SIZE):
        try:
            result = await db.actionables.insert_many(
                chunk, ordered=False  # Continue on failures
            )
            inserted_ids.extend(result.inserted_ids)
        except BulkWriteError as e:
            # Partial success - don't fail entire batch
            successful = e.details.get('insertedIds', [])
            inserted_ids.extend(successful)

            # Dead letter queue for failed docs
            failed_docs = [d for d in chunk
                          if d['_id'] not in successful]
            await dlq_service.publish(
                'task_creation_failures',
                failed_docs
            )

    return inserted_ids  # 90s → 5s (18× faster)

Organization Admin Dashboard Refactor

The Challenge

Admin dashboard was unusable with 6.2s load times. Backend made 3 separate API calls with server-side aggregation (JOIN + GROUP BY), causing cache inconsistencies and data mismatches. Teams complained about stale role counts after organizational changes.

The Solution

Complete architecture refactor: eliminated 2 redundant endpoints, moved role count aggregation to client-side (O(n) JavaScript Map operations), implemented cascading cache invalidation, and rebuilt UI with modular tab components. Single API call now provides single source of truth.

Technical Highlights

Subtask Parent Resolution System

The Challenge

Critical bug blocked 850+ disclosure operations monthly. Data model only mapped parent tasks to communications, but users created disclosures from subtasks. System threw 404 errors, causing compliance workflow failures.

The Solution

Designed parent resolution algorithm with intelligent fallback: try direct mapping (O(1) for parent tasks), then check subtask mapping collection, resolve parent ID, fetch parent's communication. Added background job for orphan detection and data integrity monitoring.

Technical Highlights