Reduced tax paperwork from 165 hours to near-zero with AI automation

Project Description:

Deductly is an AI-driven tax substantiation system that turns a short-term rental (STR) owner's Google Calendar into an IRS audit-ready tax documentation engine. By embedding structured hashtag metadata into everyday calendar events, Deductly automates the creation of time logs, mileage records, and expense tracking — eliminating the 90–165 hours STR owners spend annually on tax paperwork and unlocking an estimated $1,500–$3,000 in missed deductions.

Achievements: Built a production-grade system comprising 10 interconnected tables, 11 automations, and 9 custom interfaces — achieving 85%+ automation reliability within a no-code environment. Reduced AI operating costs by ~47% through strategic field optimization (17 AI fields down to 9) without sacrificing functionality. Receipt OCR processes vendor name, amount, date, and tax category with automated confidence scoring, vendor fuzzy-matching, and duplicate detection. The calendar integration pipeline — from Google Calendar event to fully substantiated Time Log and Mileage Log — runs autonomously on a nightly schedule with zero manual intervention.

Tools: Airtable (Team plan) as the core database and automation engine, Google Calendar API for event ingestion, Airtable AI Vision for receipt OCR and odometer scanning, custom JavaScript automations for hashtag parsing and data transformation, and a planned Phase 2 integration with Google Maps API for GPS-based mileage calculation.

Methodology: A calendar-driven, staging-table architecture ensures data integrity: events are imported into a staging layer, parsed for structured metadata (#Prop, #Cat, #Qual, #From, #To), validated, and then posted to final logs via a nightly sync. Expense processing follows a chained automation pipeline — receipt upload triggers AI extraction, vendor auto-linking via fuzzy match, default inheritance, duplicate guarding, and low-confidence flagging — all without user intervention. The system was developed in iterative phases: Phase 1 focused on core automation reliability and infrastructure cleanup; Phase 2 targets income tracking, Google Drive integration, and commercial launch.

Learnings: Field-name precision in no-code automation environments is mission-critical — a single mismatch silently breaks entire pipelines. Airtable's AI vision capabilities show meaningful inconsistency across identical configurations, reinforcing the need to design manual fallbacks for every AI-dependent process. End-to-end integration testing surfaced issues that unit-level testing consistently missed. And perhaps most importantly: the highest-leverage AI isn't always the most sophisticated — it's the AI that removes friction from workflows people already use.

My next step is building the frontend for users with Lovable.

Relevant URLs: Deductly AirTable