Inspiration: Reputation is everything, and when something falls through the cracks, it can be a significant uphill climb to get it resolved.
What it does: Autonomous reputation monitoring system that scans Google reviews, classifies credibility threats using AI, drafts professional responses, and automatically creates tasks in your CRM + alerts in Intercom. Learns from human feedback to improve response quality over time.
How we built it: Frontend: Next.js 15 + React 19 + Tailwind for real-time dashboard
AI Pipeline: Groq LLM (Llama 3.3 70B) for sentiment analysis + response drafting Integrations: Apify for Google Maps scraping, Intercom API for notifications, KafkaJS for Redpanda streaming, Sentry for monitoring Architecture: Serverless on Vercel with /tmp file persistence for cross-invocation state sharing
Challenges: we ran into Serverless state management: Lambda instances don't share memory - solved with /tmp file syncing between invocations
LLM JSON reliability: Groq occasionally returned malformed JSON - built progressive cleanup with fallback defaults Intercom contact creation: 409 conflicts on duplicate emails - extracted contact IDs from error messages Real-time polling UX: Dashboard needed to show pipeline stages smoothly - added intermediate status updates with sleep delays
Accomplishments: that we're proud of Self-improving feedback loop: Agent actually learns from operator corrections and adapts tone/approach
Production-grade integrations: Not mocked APIs - real Intercom messages, real CRM entries, real Kafka events Demo resilience: Graceful degradation when integrations fail (skips instead of crashes) Clean architecture: Scout → Analyst → Action pipeline is modular and testable
What we learned: Serverless forces you to architect for statelessness (global + /tmp hybrid worked well)
LLMs need defensive parsing - never trust raw JSON output User feedback is the easiest way to train an agent without MLOps infrastructure Visual pipeline diagrams make complex workflows immediately understandable to non-technical users
What's next for Group 7: Multi-source monitoring: Expand beyond Google to Yelp, Facebook, Trustpilot
Proactive outreach: Auto-respond to reviews (with human approval) via Google My Business API Trend detection: Alert on credibility score velocity ("you're losing 5pts/day - investigate") Response templates: Learn common patterns and suggest pre-approved response templates Competitor benchmarking: Compare your credibility score against local competitors
Log in or sign up for Devpost to join the conversation.