KICK WinWin — Chat Engine

KICK chat moves faster than any streamer can read. The hard part isn't counting messages — it's deciding what to do while the audience is still there.

Comment

Inspiration

KICK chat moves faster than any streamer can read. Viewers ask questions, repeat requests, hype up moments, react with emotes, and signal generosity through gifts or sub-gifter activity. The hard part isn't counting messages — it's deciding what to do while the audience is still there. We kept coming back to one question: what is chat trying to make happen right now, and what safe action should the streamer take next?

What it does

KICK WinWin turns live chat into a full engagement loop across three stages:

  • Before stream — a research/search view (word cloud, TL;DR summaries, gifter cues) helps the creator prepare topics and interaction seeds.
  • During stream — on /live, KICKWIN watches chat and gift-like signals in real time, detects lulls, hype spikes, repeated requests, questions, and gifter moments, then suggests a safe action (a question, a poll/quest, a celebration) — always kept behind streamer approval before it goes public. Viewers can shape the moment through Community Quest votes, gestures, and a collaborative canvas.
  • After stream — clicking End live opens an instant Stream Summary (estimated gift revenue, unique chatters, gift-driving supporters, a recap-style Event Intelligence timeline, and one best next-stream topic). The full /stream-recaps view goes deeper: Topic Map, Gift Drivers, Audience Intelligence, high-frequency questions, and a Next-Stream Strategy Brief.

How we built it

  • Frontend: Next.js 16, React 19, TypeScript — a KICK-style creator and viewer UI (/live, /stream, /research, /stream-recaps, /demo).
  • Backend: FastAPI + Pydantic for deterministic live/recap analysis, with a LangChain-ready insight service for optional Claude enrichment.
  • Real Kick integration: a signed webhook receiver (POST /webhooks/kick) verifies Kick's RSA-SHA256 signature against Kick's published public key, rejects replayed events outside a 5-minute window, de-duplicates by message_id, and persists real chat.message.sent / kicks.gifted events — optionally into Supabase, with a local JSON fallback so the demo stays reliable offline.
  • OAuth2: standard Kick login (authorization-code flow) plus a client-credentials app token used to subscribe channels to webhook events.
  • Deterministic rules keep the whole demo reliable without an LLM key; an optional AI enrichment path adds richer insight when one is configured.

Challenges we ran into

  • Keeping the audience's actual signal (questions, hype, gifting) legible inside a chat stream that moves far faster than any human can read.
  • Verifying real webhook signatures correctly (RSA-SHA256 over message_id.timestamp.raw_body) and handling Kick's redelivery/replay behavior without dropping or duplicating events.
  • Being disciplined about real vs. simulated data — every proxy metric (e.g. "gift activity proxy", "returning audience proxy") had to be labeled honestly instead of implied as ground truth.
  • Presenting a two-sided (streamer + viewer) story in a single demo window without real cross-device state sync, which we solved with a dedicated /demo split-screen page sharing one React state tree.

What we learned

  • Webhooks and OAuth solve two different problems (event delivery vs. authorization) and shouldn't be conflated — but both matter for a real platform integration.
  • A hackathon demo is stronger when every simulated mechanic is clearly labeled rather than implied to be live multi-user data.
  • Small, fast, honestly-scoped features (a 3-second gesture prediction window, a fast collaborative canvas reveal) demo better under a 5-minute clock than slower, more "realistic" versions of the same idea.

Accomplishments that we're proud of

A real, signature-verified Kick webhook pipeline feeding a full before/ during/after engagement loop — not just a mockup, but a system that can actually receive and act on live Kick chat data.

What's next for KICK WinWin

Wire the pre-stream research view directly into the live signal detector, extend gift-driven topic detection with more granular Event Intelligence drilldowns, and move from local JSON fallback to Supabase as the default store for production reliability.

Built With

Share this project:

Updates