Dark is a user feedback analysis engine that automatically surfaces sentiment patterns, feature requests, and churn signals from marketing team feedback sources (surveys, support tickets, reviews, social media). It enables marketing teams to prioritize campaigns and messaging based on real user sentiment rather than intuition, reducing wasted spend on misaligned messaging and accelerating time-to-insight from days to minutes.
Marketing teams currently spend 15-20 hours per week manually reading and categorizing customer feedback across fragmented sources (Intercom, Zendesk, App Store reviews, Twitter mentions, survey responses). This creates three friction points: (1) only high-volume feedback surfaces while edge cases and small-but-vocal segments are missed, (2) sentiment analysis is inconsistent across team members, leading to conflicting campaign hypotheses, and (3) insights arrive too late to influence sprint planning or campaign creative. Teams report making campaign decisions based on 3-5 recent conversations rather than comprehensive data, resulting in misaligned messaging that underperforms by 20-30% vs. data-driven cohorts.
As a marketing manager, I want to ingest feedback from multiple sources (Zendesk, Intercom, App Store reviews, Slack) into a single dashboard so that I can see the complete feedback picture without switching between tools.
As a campaign strategist, I want Dark to automatically extract and cluster feature requests by frequency and urgency sentiment so that I can prioritize which product benefits to highlight in next month's campaign.
As a performance marketer, I want to see real-time churn signals and negative sentiment spikes by cohort so that I can pause underperforming campaigns before budget is wasted.
As a social listening manager, I want Dark to identify which competitor features or messaging are mentioned positively in user feedback so that I can brief product on competitive threats or opportunities.
As a content marketer, I want to access anonymized customer quotes that match specific sentiment/topic filters so that I can find authentic testimonials for case studies without manually searching support tickets.
Adoption: ≥60% of marketing team (min. 5 users) logs in weekly within 60 days of launch, measured via product analytics session tracking. Insight velocity: Average time from feedback ingestion to campaign team viewing relevant insight ≤8 hours, measured via timestamp comparison (feedback received → dashboard view). Decision impact: ≥40% of marketing campaigns launched post-Dark launch cite at least one insight from Dark in their brief, measured via campaign retrospectives and tagging. Feedback coverage: Dark surfaces ≥85% of feedback mentions for top 20 topics within a 7-day window (validated monthly via manual spot-check against raw feedback sources). False positive rate: NLP-generated sentiment and topic labels have ≥90% accuracy against human review of random 200-sample validation set, measured bi-weekly.
Duplicate feedback handling: Same support ticket forwarded via multiple channels (email + Slack) or auto-forwarded between tools could create false duplicates; must detect via content hash + source timestamp within 1-hour window. Language and encoding: Non-English feedback (Spanish, Chinese, French) must not crash sentiment classifier; system shall gracefully degrade and label as "unclassified language" rather than return false English sentiment. Missing or inconsistent metadata: Zendesk tickets may lack customer segment/company info; system shall not crash on missing cohort data but display as "Unassigned" segment separately. API rate limiting and quota exhaustion: Zendesk, Intercom may rate-limit mid-sync; system shall implement exponential backoff and queue failed ingestions for retry; alert admins if connector is consistently hitting limits. Timezone handling: Feedback timestamps from global sources may be in UTC, customer local, or unspecified; system shall normalize all to UTC for consistency and display user's local time in UI with timezone label. Sentiment classifier drift: NLP models trained on historical data may misclassify emerging slang or product terminology; low-confidence predictions (<70%) shall always show to users so they can provide feedback. Large feedback text: Support ticket transcripts or user interviews may be >10K characters; system shall truncate preview to 500 chars in UI but allow full text view; clustering must handle long documents without memory overflow. Connector authentication expiry: OAuth tokens expire; system shall handle token refresh automatically and alert admins if refresh fails; failed auth shall not ingest new feedback but shall not delete historical data. Null or sparse data scenarios: New customers with <100 feedback items in first week; dashboard shall not crash but display "Insufficient data for clustering" message and suggest waiting. PII exposure: Feedback may contain customer names, email addresses, or sensitive account info; system shall not store PII separately but shall warn users not to export sensitive data and provide anonymization toggle.
⚠ Which 3-5 feedback sources should be MVP vs. roadmap? (e.g., is Slack channel integration critical or Phase 2?) ⚠ Should Dark support custom CRM integrations for cohort data, or only assume Zendesk/Intercom segmentation tags? What sentiment classification model should we use—existing third-party API (e.g., AWS Comprehend, OpenAI API) or fine-tuned custom model? (Cost vs. accuracy tradeoff) What is the acceptable latency for alert delivery—must alerts fire within 5 minutes of feedback ingestion, or is hourly digest acceptable? Should topic clustering be fully automated (unsupervised) or allow marketing team to manually define/label topics? How do we handle feedback from customers who have churned or are no longer in CRM—should churn signal detection only apply to active cohorts? What's the data retention and compliance requirement—GDPR/CCPA compliance for storing customer feedback? Should we support feedback in non-English languages in MVP, or English-only initially with i18n roadmap?
Zendesk API (OAuth + ticket export), Intercom API (OAuth + conversation export), App Store Review API (requires Apple credentials; may need third-party scraper service if direct API unavailable). NLP/ML inference service: Either third-party API (AWS Comprehend, Google Cloud NLP, OpenAI) or internal ML pipeline with embedding model deployment. CRM data sync (optional MVP): If cohort-based analysis required, integration with Salesforce, HubSpot, or native Zendesk/Intercom user fields. Slack API for alert delivery (incoming webhooks). Authentication/SSO: Marketing team login via existing company identity provider (Okta, Azure AD, or email-only MVP). Analytics service: Event tracking for adoption metrics (Segment, Mixpanel, or internal event sink). Data warehouse or append-only event log: To store raw feedback and allow historical querying without re-ingesting from source APIs. Deployment infrastructure: Containerized service capable of handling concurrent API syncs and NLP inference (Kubernetes or Vercel/Heroku acceptable for MVP).