Inspiration Every time Mccan growth team missed a breakout hashtag on X by even 1 day, engagement on the follow-up post dropped >40 %, creates client frustration. We wanted a way to see spikes before they hit everyone’s timeline, translate raw buzz into ready-to-post angles, and let non-technical marketers steer the system without waiting on dev cycles.
What it does Realtime Trend Detector ingests X/Twitter trends every 60 seconds, enriches them with live stream stats, scores each topic on five signals (rank, volume, velocity z, engagement, brand-similarity) and fires tiered alarms:
Red (“Post-NOW”) → Slack ping + auto-drafted hook & CTA Orange (“Opportunity”) → queued for same-day content Yellow (“Seed”) → logged for newsletters/podcasts A web dashboard shows the feed, a velocity/time-map radar, and editing panels where anyone can add Countries or tweak scoring weights in real time—no code deploy.
How we built it
- X Trends API exports tweet_volume and trend text (ie hashtags) into a timestamped Supabase table
- WOEIDs are exposed on the same schema on a whitelist
- MCP exposes the schema as resources and tools to track volume, trend text, and WOEIDs for functions answering questions such as "What trends are in X location?", "What are the top trends over the past 24 hours?", and other permutations.
- AI Agents on the frontend that 1. predict upcoming trends due to tweet volume/time, and 2. predict if a postulated topic will be a fit for trends being detected.
Challenges we ran into API rate limits – had to rotate bearer tokens and add smart back-off. Noise vs. signal – early models flagged celebrity gossip; cosine brand-similarity fixed 80 % of false positives. Latency – embedding calls were bottleneck; added Redis + batched requests to cut cycle to 7 s. UX for speed – designers shaved every click; single-key “Handled” shortcut dropped T-to-post by 22 %.
Accomplishments that we're proud of Median time from alarm to published post 3 m 42 s (Red tier). 85 % precision: Red alarms drive ≥5 % engagement uplift vs. baseline. Non-tech teammates now add locations/weights themselves—0 SQL tickets last month. Plugged directly into Parsed’s content grid, so hooks propagate to Instagram/TikTok planners with no copy-paste.
What we learned Velocity spikes alone are noisy; combining engagement ratio + similarity is crucial. Marketers trust the system only when they see transparent feature breakdowns (“why hot” chips). Async architecture with tiny, purpose-built functions beats monolith for real-time constraints.
What's next for Parsed Cross-platform detectors – TikTok, Reddit, and LinkedIn. Predictive module – Monte Carlo sims to flag incoming trends 30 min early. Auto-publish (opt-in) – generate + schedule posts directly for low-stakes channels. Personal Brand Module - Relate concepts to leverage personal brand for CEOs
Test account (Trends tab) https://app.beparsed.com user: [email protected] pass: Parsed1
Log in or sign up for Devpost to join the conversation.