Recent accomplishments

1. Bench Accounting Team Automate

   Solely responsible for designing and implementing a backend service that enriches transactions with merchant details using LLMs and a retrieval-augmented generations architecture. By surfacing this information in our internal tools using micro-frontends, saved each human specialist 2.4 hours/day of work and, as a result, around $14,500 per month.

2. Developed LLM-based categorization assistant to guide customers through categorizing according to tax rules, which auto-categorized 230,000 transactions using customer conversations within 3 months, equivalent to ~30,000 months of human bookkeeping or 37% of the work done by specialized human teams in the same period.

3. Wrote a Python backend using Lambda, API Gateway, and React micro-frontends to surface groups of similar uncategorized transactions to customers, enabling them to auto-categorize in groups. This feature increased the manual categorization rate by 55% upon deployment, saving the business the equivalent of ~110 workdays per month in manual human bookkeeping and $22,600 per month.

4. Trained and deployed per-client models to categorize transactions that other categorization models cannot recognize. Raised the auto-categorization rate by 30%, with 98% accuracy, saving the business ~53 days of work per month previously performed by human bookkeepers within two months, equivalent to a cost saving of $10,500 monthly.