211 Commits, 0 New Features — The Week That Actually Mattered
Hey IH! Week 4 update on Underpriced AI — the AI pricing tool for resellers.
Quick Recap
Last week I shared that a customer complaint led me to discover our AI costs were 10x what we thought ($0.52/scan, not $0.05). That forced a full audit of the billing code, which turned into a full audit of... everything.
211 commits later, here's the honest breakdown:
~40% was mobile bug fixes (overflow, layout, safe areas)
~20% was anti-fraud and email verification
~15% was admin tooling (partnership CRM, forecasts, analytics)
~15% was cost optimization and pricing changes
~10% was actual user-facing features
The biggest impact came from the boring 40%.
The Mobile Wake-Up Call
I'd been testing on my iPhone and a Pixel. Things looked fine. Then I started checking on Samsung Galaxy A-series phones (the most popular Android phones worldwide) and found the app was borderline unusable:
Item titles running off cards
Descriptions pushing the page sideways
Bottom nav bar covering buttons
Stats cards breaking on narrow screens
20+ commits of overflow fixes, safe area insets, and responsive adjustments. Zero of these were reported by users. Users don't report layout bugs — they just churn.
Lesson: If your target market includes people who sell $20 items from thrift stores, they're not using $1,000 phones. Test on cheap devices.
Free Tier Abuse Was Real
Found users creating 3-4 accounts with different email addresses to get unlimited free scans. Same IP, same device fingerprint.
Built an anti-fraud system:
Email verification required before scanning
IP-based duplicate detection with confidence scoring
Soft-ban system with self-service recovery
50+ disposable email domain blocklist
Merge tool for legitimate duplicate accounts (Google vs email signup)
This was a full week of work that generates zero new revenue. But it stops the bleeding on free-tier costs.
Features That Actually Shipped
Auto-save photos to device. When users scan an item, the photo now downloads to their phone automatically. Sounds obvious, but before this, photos were only saved to our servers. Users had to retake the photo when listing on eBay/Poshmark. Small fix, big workflow improvement.
Status-tiered inventory sorting. Items now sort: Analyzed (needs action) → Listed (actively selling) → Sold (completed) → Archived. Before, everything was mixed together sorted by date. Now your "to-do" items are always at the top.
"Death Pile" rename. Renamed "List Next" to "Death Pile" — that's what resellers actually call their unsorted inventory. Know your audience's language.
CSV export for inventory. Users can export their full item inventory (not just scans) as CSV for spreadsheet tracking.
Partnership outreach CRM. Built an admin tool to discover estate sale companies via AI, enrich them with contact info, and send outreach emails. This is my B2B acquisition channel — estate sale companies recommend our app to their customers.
Cost Optimization (following up on last week's $0.52/scan discovery)
Trimmed Claude system prompt 21% (4,200 → 3,300 tokens)
Reduced web searches from 3 → 1 when users provide hints
Skip Perplexity enrichment on re-scans (item already identified)
Fixed credit-pack-only users getting 30 free subscription scans
Per-scan cost tracking now captures full token breakdown (input + output + cache)
Raised credit pack prices so subscriptions are always cheaper per-scan
Removed the 1-scan pack (negative margin after Stripe's $0.30 fixed fee)
Result: Cost per scan trending ~30-40% lower than the $0.52 we discovered last week.
Pricing Changes
Did a full pricing overhaul:
Repositioned credit packs as premium (always more expensive per-scan than subscriptions)
Removed the 1-scan pack entirely — after Stripe's $0.30 fixed fee, I was losing money
Added lifetime scan counter per user for analytics
What's Next
Acquisition is still the bottleneck. The product is solid, mobile works properly now, fraud is under control. Focusing on:
Partnership outreach to estate sale and auction companies
Content marketing (SEO blog posts, building in public)
Potentially short-form video showing the scan-to-sell workflow
Questions for IH
Anyone dealt with free-tier abuse at scale? My confidence-scored IP grouping works, but I'm curious what others have tried.
What's your best B2B cold outreach channel? Email? LinkedIn? Something else?
Building in public at underpricedai.com
