TrendGen: AI Trend Intelligence Platform
An end-to-end trend intelligence platform that mines social conversations, discovers emerging consumer trends with transformer-based topic modelling, names and scores them with an LLM, and forecasts which signals will break out next, all driven from a real-time admin console.
- Posts processed per run
- 100K+ Posts processed per run
- Emerging-signal hit-rate
- 74% Emerging-signal hit-rate
- Avg forecast lead-time
- ~17 days Avg forecast lead-time
Project Overview
An end-to-end trend intelligence platform that mines social conversations, discovers emerging consumer trends with transformer-based topic modelling, names and scores them with an LLM, and forecasts which signals will break out next, all driven from a real-time admin console.
The Problem
Teams needed to spot emerging consumer trends early from noisy social data, but manually analysing hundreds of thousands of posts was slow, inconsistent, and impossible to act on in time.
The Solution
A production-shaped ML platform where raw posts flow through an embedding, clustering, and LLM-naming pipeline into ranked trends, while a Qdrant + DTW forecasting layer surfaces emerging signals early, all orchestrated as async jobs behind a real-time React admin console.
Key Features
Generation engine
Guided, tabbed workflow to train a BERTopic model, generate trends over a filtered post set, or segment them category-wise.
Trend explorer
Ranked, searchable, filterable trend cards with confidence scores, keywords, tags, and live post volume.
Trend deep-dive
Per-trend detail with growth rate, language mix, category breakdown, and the real posts driving each signal.
Emerging-signal forecasting
Qdrant vector search plus DTW time-series matching scores early signals before they peak, at a 74% hit-rate.
Async job orchestration
Long-running training and generation jobs run with per-job status, stage, and progress streamed live to the UI.
Model registry
Trained BERTopic models are versioned in Azure Blob Storage so prediction runs stay fast and repeatable.
Architecture & Flow
Live Build
What I Delivered
- Unsupervised NLP pipeline: embeddings, BERTopic clustering, and LLM-assisted trend naming
- Emerging-signal forecasting with a Qdrant vector DB and DTW time-series matching
- Async job architecture with SSE-streamed progress for heavy ML runs
- React + TypeScript admin console covering the full trend lifecycle across nine surfaces
- Versioned BERTopic model registry backed by Azure Blob Storage
- Scheduled weekly pipelines and daily feedback evaluation via APScheduler
Ready to build something useful?
Tell me what you're working on - I can help you plan, design, build, and launch an AI-powered product, automation system, web app, or mobile app.