Conference speakers prepare talks weeks in advance for an audience they've never seen — the first real feedback arrives during Q&A, when it's too late to adapt. CrowdShift closes that gap by turning registration data, which already exists in every ticketing system, into live audience intelligence.

As attendees register, CrowdShift snapshots the audience at capacity milestones (25/50/75/100%) and tracks how it evolves: a talk that starts senior-heavy can be 55% junior developers by sold-out. A timeline slider lets speakers scrub through these snapshots and watch the demographics shift in animated charts. At any point, an AI agent (Gemini) reads the current snapshot and generates a structured brief — who's in the room, what changed since the last milestone, and concrete recommendations like which comparisons to lead with and which deep-dives to cut.

The product serves three roles with strictly separated data access. Speakers see aggregate demographics and pre-submitted, upvotable attendee questions for their own talk. Organizers get the full 500-row attendee grid with live filtering. Sponsors see anonymized cohorts only — segments under five people are suppressed (k-anonymity), and the separation is enforced server-side with role-scoped endpoints and token auth, not hidden in the UI. Attendees participate too: with just a ticket ID they can submit questions to speakers before the event, creating a feedback loop between the room and the stage before anyone walks on it.

Built in 24 hours for the Progress x GitNation hackathon: React + TypeScript + Vite frontend on KendoReact (Grid, Charts, Slider, and 10+ components, fully re-themed with a custom dark/light design system), FastAPI + SQLite backend with 500 seeded attendees whose registration behavior encodes realistic demographic waves, and a brief-generation pipeline with structured JSON output, caching, and a rules-based fallback so the demo can't fail on a flaky API call.

FIP (feline infectious peritonitis) treatment is a marathon — often 84 days plus a 12-week relapse watch — and the thing a caregiver needs most is to see the trend over days, not guess from memory at 2am. FIP Tracker turns a simple daily log into clear charts (weight gain, the dry-FIP fever cycle, A:G recovery) so "is he actually better than last week?" finally has an answer.

A three-phase model (monitoring → treatment → observation) with phase-aware guidance grounded in the UC Davis / Pedersen GS-441524 protocols. The daily log covers weight, temperature (fever auto-flagged), appetite/energy/interest scores, litter box, symptoms and medication; a bloodwork section auto-computes the A:G ratio and charts the markers that track recovery. All charts are hand-built SVG — no chart library.

Full auth (Google/Facebook OAuth), persistent storage, a printable one-page vet report (save as PDF), read-only share links, and shared cat co-ownership so a couple can track the same cat on separate accounts via a join-code-and-approval flow. Fully internationalized (English + Polish) with light/dark theming.

Self-hostable and privacy-first: no ads, no data selling, research-data sharing strictly opt-in. Uses Node's built-in node:sqlite (zero native dependencies), Dockerized behind Caddy with automatic TLS, plus a security pass — Content-Security-Policy and hardening headers, per-IP rate limiting on auth endpoints, and stateless signed-cookie sessions.

GitHub Actions parse project metadata from repo files. Medium's RSS feeds article data in automatically.

A daily CI/CD job triggers full rebuilds — and so does every git push. The entire stack is stateless and declarative: no database, no manual updates, just automated pulls and static generation.

Deployed on AWS using S3 + CloudFront + Lambda, staying within free tier limits. Even the contact form runs through Lambda — no backend server, just functions doing exactly what's needed and nothing more.

You're looking at the output.