introduction

Blueberry Social

E-commerce brands still lose sales in Instagram comments. High-intent questions sink between spam, and teams ignore AI when they cannot see or fix its choices. Blueberry turns those comments into revenue with on-voice replies, edits that teach, and automation for routine asks. I led visual design and prototyping, shipped the inbox, brand voice, and automation surfaces, and co-built launch prototypes through a mid-project pivot.

overview

Reply faster, scale safer

The redesign focused on three essentials. A hybrid inbox balances scan speed with context. A brand voice flow learns from real edits. Automation rules keep controls clear and reviews simple. Together they turn comment threads into on-brand revenue moments without risking trust.

problem

Unanswered buyers, lost revenue

A single drop can spark five hundred comments in a day. Buyers asking for links or shipping answers get buried under updates and spam. Teams hire more people, reply late, or refuse AI because one public miss can bruise the brand.

exploration — inbox

Hybrid inbox: scan fast, reply with context

Showing every comment at once overloaded operators, while one-at-a-time views slowed them down. The final hybrid inbox keeps rows collapsed for scanning, then expands to reveal the thread, AI draft, and source context in one place. Operators move from scan to send without tab hopping.

exploration — brand voice

Exploration 1: Real-time preview split attention

A split layout with guidelines on the left and live tests on the right promised instant feedback yet forced constant context switching. People rewrote traits, checked output, and looped, unsure which pane to watch.

Exploration 2: Progressive disclosure worked

Collapsing setup into a single column let teams define tone without distraction. A modal preview showed multiple cases side by side, exposing inconsistencies the split view hid. The clearer flow encouraged deeper, calmer guideline work.

exploration — automation

Exploration 1: Rules that read like sentences

Automation logic now reads the way people think: “When someone mentions your brand, then reply.” Plain language with dropdowns blocks vague triggers and mirrors familiar tools like Apple Shortcuts, so non-technical teams know what fires.

Exploration 2: Make complex rules safe

Early modals cramped decisions and hid review. A dedicated builder page gives room to reason through complex triggers, adds a review step before launch, and logs every change so automation can scale without public mistakes.

final design

Onboarding without friction

Paste a store URL and Blueberry pulls the brand name, icon, and connects Meta Business Suite. Teams land inside with their identity already set, so the first reply arrives before friction creeps in.

Flexible, fast inbox

The hybrid inbox keeps the queue light for scanning, expands for context, and adds quick filters across platform and comment type. Inline assist handles rewrite or regenerate before send, making AI help visible and controllable.

Brand voice learned from real edits

Teams define traits in a focused view, preview across scenarios, and pull knowledge from their site to ground responses in real policies. Each edit shapes future drafts, so the AI mirrors tone without long training loops.

Automation with guardrails

A dedicated page lets operators draft sentence-based rules, review the “when → then” logic, and launch with human-in-the-loop defaults. Confidence gates, audit logs, and one-click revert keep experiments safe.

retrospective

Prototype to learn, not to impress

Scrappy prototypes surfaced blockers—hidden AI and unclear next steps—days sooner than polished visuals. Next time I would instrument enable-in-first-session, edit distance, and first-reply time even earlier.

Bring engineering in early

Weekly feasibility checks turned safety ideas into shippable constraints. Moving from modals to full pages, setting confidence thresholds, and adding logging all landed because engineering stayed in the room.

Cut scope to raise clarity

Focusing on comments instead of DMs sharpened the hybrid inbox and the learning cues. Future sprints will time-box side bets and end them fast when they dilute the core story.

Design through pivots

The mid-sprint pivot reinforced the value of tackling the riskiest assumptions first. When the market changed, I narrowed in on the blockers research exposed and iterated directly against them.