InsighT

About the project

We set out to give PMs the power to see, understand, and act. T‑Insight stitches together fragmented signals into a single intelligence layer: what’s happening, where it’s happening, how bad it is, and what to do next. We built a scalable ingestion pipeline, a thoughtful UI that emphasizes clarity, and an AI layer that transforms insights into prioritized opportunities. Along the way, we solved SSR pitfalls, map layering/z‑index issues, and model compatibility quirks. This helped deliver a focused, production‑ready experience that PMs can trust.

Inspiration

T‑Insight was born from a simple question PMs ask daily: “Where are customers hurting right now, and what should we do about it?” Feedback is scattered across social media, outage trackers, forums, and support logs. We wanted a single pane of glass that turns raw noise into prioritized, location‑aware, actionable intelligence.

What it does

Consolidates multi‑source customer signals (Reddit, outage trackers, community forums, internal feedback) into a unified timeline.
Computes a Customer Happiness Index (CHI) and trends by product area (Network, Mobile App, Billing, Home Internet).
Visualizes geographic clusters via a Leaflet heatmap and overlays live outage markers.
Surfaces “Top Emerging Issues” ranked by intensity and sentiment.
Generates AI‑powered, actionable opportunities with priority/urgency, root cause analysis, recommended actions, impact, and stakeholders using Google Gemini 2.0 Flash.
Provides a clean, on‑brand dashboard with charts (Recharts), timelines, distribution, source breakdown, and an early‑warning panel.

How we built it

Frontend: Next.js (App Router), React, TypeScript, Tailwind, Radix UI.
Data: Supabase (Postgres, RLS), server actions, SSR/edge‑friendly API routes.
Ingestion: Scrapers for Reddit, News, DownDetector, IsTheServiceDown, Outage.Report, plus customer feedback; processed into normalized “signals”.
Processing: Sentiment scoring, topic detection, deduplication/merging, and intensity weighting.
Geo: Leaflet + react‑leaflet + leaflet.heat with OpenStreetMap tiles and Nominatim geocoding.
Visualizations: Recharts for CHI, timelines, distribution; custom UI for issue velocity and early warnings.
AI: Google Gemini 2.0 Flash for insights on emerging issues via secure server‑side API routes.
Ops: CRON‑secured ingestion endpoint; environment variables for Supabase and Gemini.

Challenges we ran into

SSR vs client‑only libraries: Leaflet requires the browser. Fixed with dynamic imports and a client wrapper.
Hydration and z‑index conflicts: Resolved dialog/modal stacking over a sticky navbar; eliminated <div> inside <p> hydration issues.
Model/version mismatches: Updated Gemini to 2.0 Flash and hardened error handling for model availability.
Data sparsity and weighting: Tuned intensity/sentiment normalization so heatmaps show meaningful medium/high hotspots (not only “low blue”).
Time windows: Standardized 24‑hour windows with graceful fallbacks for empty datasets and added diagnostics.

Accomplishments that we're proud of

A cohesive, brand‑aligned PM dashboard that goes from raw signals to actionable opportunities.
Real‑time, geo‑aware heatmaps and outage overlays using only free mapping/geocoding services.
AI‑generated, non‑generic recommendations with priority/urgency and stakeholders—useful out of the box.
Robust ingestion + processing pipeline with deduplication and topic detection.
Clean, accessible UI with responsive layouts and clear information hierarchy.

What we learned

Designing for SSR + client‑only libs requires disciplined boundaries and dynamic loading.
Weighting and normalization are critical for trustworthy visual signals (especially heatmaps).
Good AI prompts and response validation dramatically improve actionability.
Small DX investments (diagnostic logs, fallbacks, typed schemas) accelerate debugging and iteration.

What’s next for InsighT

Real‑time streaming updates via websockets for live dashboards.
Advanced filters (time range, geo bounding boxes, product area, channel).
Deeper ML: trend forecasting, anomaly detection, and clustering of root causes.
Workflow integrations (Jira/Linear/Slack) to turn insights into tracked opportunities.
Team collaboration: comments, assignments, and change history.
Mobile‑friendly PM views and alerting for critical spikes.