Recoverflow

The day after submitting to the hackathon I finally had a moment to look back at the whole system. Turned out something surprising — across those 7 days we spent more time thinking about architecture than writing code. Days 9-12 were 4 solid days of talking to my mom, my aunt about real collections scenarios, listing D-001 through D-022 design decisions, drawing drawio v1/v2/v3. Days 13-15 were the only code days. This post is the full design journey — why I picked ‘outstanding balance’ over ‘gross invoice,’ why 9 agents in 3 layers, the 4 hidden cases nobody talks about — all from imagining real conversations with real people on bad days.

Day 2 was Voice Agent day. I wanted a ‘real two-way conversation’ — Sarah should hear the customer say ’next week,’ then push back to ask ’next week which day, and how?’ To make sure she actually worked, I called myself. And immediately hit a bug — I said ‘I need to ask my boss, I’ll email you in 2 days,’ and Sarah treated it as evasion and escalated. I was stunned. Added Phase 3b. Then ran 9 scenarios — all pass. This post is the full story.

I built Recoverflow at a hackathon — an AI collection system for small businesses with cross-border receivables. I found the gap in the middle market: lawyers only take big cases, and collection agencies run every case through the same blunt SOP. The first agent, ‘Pre-flight,’ extracts the contract clauses and routes by outstanding balance (not gross invoice) into three paths: under $3K runs Lite, $3K–$40K runs the full In-Spot pipeline, over $40K we don’t run the system at all and refer you straight to an attorney with a vetted list.