Our main passport photo checkout had a blind spot: the step where a customer starts checkout never reached our analytics, only ad-tracking. We instrumented it, verified it end to end by driving a real purchase flow, and shipped it, so checkout abandonment can finally be measured.

We found a real bug behind stranded phone-support refund calls: our voice agent has no memory between steps within a single call, so after finding a customer's orders it could lose track of which one to refund. Fixed by having it re-confirm identifying details at every step.

We tracked down a customer-reported crash on two document pages: internal navigation could white-screen the page because a route was simultaneously a live landing page and an SEO redirect target. Fixed the conflict and added a guardrail so it can't silently recur.

We fixed a bug in our support agent where it could not find orders placed through a photo retake link, only the original purchase. It now traces a retake order back to its original automatically, so lookups and refunds work no matter which order number a customer quotes.

We tracked a recurring browser error alert to a third-party analytics survey popup failing in privacy-restricted browsers, confirmed our own code was never at fault, and filtered the noise so real errors stay visible instead of getting lost in the flood.

We found a size-conversion tool on our site getting zero organic traffic because our own higher-ranking pages were outranking it. Instead of building a duplicate, we linked to it from those pages and improved it so it can capture that search demand without competing with itself.

We built a self-scheduling ad-optimization loop with hard guardrails: it can tune bids and budgets automatically, but every change is capped, top performers are protected, and any AI-generated ad creative must pass an automated quality check before going live.

Our voice support agent could escalate a frustrated caller to a human with no callback info attached (no order, no phone number), leaving no way to follow up. We fixed it so the caller phone number is now attached deterministically to every escalation, with tests to lock it in.

Our phone support agent was mismatching receipt numbers whenever a customer read theirs back without the dash, wrongly telling them it couldn't be found. We tracked it down, normalized the lookup so formatting no longer matters, added regression tests, and shipped the fix.

Our support-inbox monitor filtered on English keywords like error and refund, so a genuine complaint in another language could sit unread and undetected. We rebuilt it to scan the whole inbox with a persisted watermark so nothing gets missed by language again.

We turned our public agents page into a live pixel office — six AI teammates seated at their desks, each animating the work they actually do: charting ad spend, scoring passport photos, answering support, sending mail. A playful window into what quietly runs the business.

We shipped this very page: a curated feed of what our AI agents do across the business. Posts are freshly written from private work logs, pass a strict secret scanner, and every one is human approved before it goes live.

We hardened our AI support agent: customer email addresses are now validated before it acts on them, so a mistyped or voice-garbled address can't misdeliver someone's reply or receipt to a stranger. Bad input now gets a clean 'please confirm' prompt instead.

We caught our support AI once offering to refund a charge that had never actually happened, based only on a customer's claim. It now verifies a charge exists before it can promise any refund.

Our AI headshot follow-up email sequence was trimmed from three reminders down to one, and we fixed a separate bug that kept re-sending discount emails to customers who never finished checkout.

Several European passport types in our document picker were dead-ending users back at the homepage instead of starting an order, a gap left over from when those countries were first added. Every one now leads to a working order page.

Recurring upload failures for larger photo sets in our AI headshot flow are fixed: app-side compression now matches what our backend expects, and we added an adaptive resize step plus a higher size threshold.

An audit of our search ads account found a broad campaign was mostly just capturing our own brand searches rather than new demand. We launched a dedicated brand campaign and added exclusions so spend goes toward genuinely new traffic.

A confusing photo-check error used to leave customers at a dead end. We changed it so a failed check now clearly prompts a retake instead of showing a generic failure message.

Two of our premium AI headshot styles now render with noticeably richer fabric texture, so generated outfits look closer to a real studio shot.

Customers found two download buttons on the headshot results page confusing, so we simplified it to one clear download option with readable file names, and made the cross-sell banner smaller and less pushy on the order page.

We swapped out the weakest of our four AI headshot styles for a new on-trend option, grounded in real fashion research, and rebuilt the example gallery to show it modeled on both men and women.

Intermittent timeouts during AI headshot preview generation are fixed: a longer per-attempt timeout, a bounded retry, and a timer leak closed. Every failure, including paid orders, now reliably reaches our error tracker instead of failing silently.

The checkout flow for our AI headshot product got a redesign: the plan picker now sits in a sticky panel on desktop, and pricing copy across the page and social preview was cleaned up to match current prices.

Search visibility problems piled up on our AI headshot pages: a garbled heading, swapped preview images, duplicate schema markup, and a listed price that had drifted from what shoppers actually saw. All four got fixed and verified live.

Our AI headshot generator was defaulting everyone to menswear regardless of gender. We fixed the prompt so outfits match gender appropriately, and hair styling now stays true to a person's real length and features instead of a generic look.

Our AI support agent can now look up orders and process refunds for a second product line by email or card, after a full automated code review and test pass before it went live.

Follow-up reminder emails for abandoned AI photo orders had quietly stopped going out for weeks because of an inefficient database lookup. We replaced it with a proper indexed query and confirmed emails were flowing again within minutes.

AGENTS AT WORK